Surgical Tool Datasets for Machine Learning Research: A Survey

This paper is a comprehensive survey of datasets for surgical tool detection and related surgical data science and machine learning techniques and algorithms. The survey offers a high level perspective of current research in this area, analyses the taxonomy of approaches adopted by researchers using surgical tool datasets, and addresses key areas of research, such as the datasets used, evaluation metrics applied and deep learning techniques utilised. Our presentation and taxonomy provides a framework that facilitates greater understanding of current work, and highlights the challenges and opportunities for further innovative and useful research.


Introduction
There are fourteen surgical specialities recognised by the American College of Surgeons, ranging from orthopaedic surgery through to vascular surgery (ACS, 2021). Each speciality has its own procedures and its own sets of surgical tools, including instruments, implants and screws designed for specific parts of the body, and for specific procedures. Rapid advances in minimally invasive surgery have led to new classifications of robotic or laparoscopic surgery and open surgery (Bhatt et al., 2018), and also to new types of instruments being introduced at a constant rate (Fig. 1).
Consequently, there are many thousands of different types of surgical tool in circulation within a hospital. Stockert and Langerman (2014) reported that just one institution processed over 100,000 surgical trays and 2.6 million surgical tools annually. There were on average 38 surgical instruments per tray, with around 6 trays used for each surgery (Mhlaba et al., 2015). Handling this volume manually in real Communicated by Jan Kybic. time and under difficult, mission-critical conditions is a challenging task requiring highly trained surgical technicians. Automating surgical tool detection and recognition through computer vision and machine learning has numerous practical applications therefore, and these applications can lead to improved efficiencies and/or reduced costs. Applications include robotic and computer-assisted surgery (Sarikaya et al., 2017;Zhao et al., 2019a), instrument position recognition in minimal invasive surgery (Zhao et al., 2017), pose recognition in surgical training (Leppanen et al., 2018;Jo et al., 2019), and instrument tracking in hospital inventory management (Ahmadi et al., 2018). Ward et al. (2021b) discussed the application of computer vision and deep learning to surgery, specifically for the identification of surgical phases and instruments in multiple surgery procedures. van Amsterdam et al. (2021) reviewed methods for automatic recognition of fine-grained gestures in robotic surgery, and highlighted the promising results obtained by deep learning based models. Garrow et al. (2021) provided an overview of deep learning models utilized for automated surgical phase recognition using data inputs such as videos or surgical instrument use, and found that laparoscopic cholecystectomy was the most common operation evaluated. Yang et al. (2020) presented a review of the literature regarding image-based laparoscopic tool detection and tracking using convolutional neural networks (CNNs), including a discussion of available datasets and CNN-based detection and tracking methods. They also presented a quantitative estimation of several performance measures. Our survey maintains a focus on surgical tools, reviews image based surgical tool detection, and provides an overview of instrument related surgical data science and machine learning techniques and algorithms. It is comprehensive in nature, covering the range of relevant research conducted in our specified time period-which was from 2015 till 2022. In particular, we maintain a focus on surgical tool datasets and on gaps in the research or on open research questions.
In this survey, we address three research questions: 1. What surgical tool datasets are used in machine learning research? 2. What machine learning methods are used in the research? 3. What are the gaps in surgical tool datasets and associated machine learning research?
Our objective, therefore, is to build a comprehensive knowledge hierarchy of applied research in surgical tool detection, classification and segmentation to guide future work. A concrete outcome is an integrated taxonomy of the methods used across the tasks undertaken in the research. We evaluate the pros and cons of each method or set of methods used in each paper, and address what is missing in the research to date. Gaps not just in the research but also in the publicly available datasets are discussed. We provide a comprehensive survey of the various datasets associated with surgical tool detection (Tables 1, 5, and 6). We address the specific challenges faced in this task and evaluate how they have been addressed. Finally, we make recommendations based on the results of the survey to encourage further work in this area.

Survey Methodology
As a logical starting point and following the approach used in similar survey work (Egger et al., 2020;Litjens et al., 2017), we rely on both PubMed and Google Scholar to conduct an initial search for literature. We chose PubMed because of its medical focus and Google Scholar because it indexes a range of peer reviewed international journals and conferences across disciplines. We expected that this strategy would provide a broader range of articles than reliance on academic databases. We used keywords to search the databasesan example search could include the keyword {"Surgical" OR "Surgery"} together with the keywords {"tool" OR "instrument"} AND {"detection" OR "classification"} AND {"deep learning OR machine learning"}. Comprehensive combinations of key words were used to ensure diligence in our search. Our reliance on Google Scholar proved to be a good strategy to develop an acceptable starting set of literature which avoided bias or preference towards any specific publisher. We also conducted other complimentary searches, such as reviewing reference lists, searching through conference proceedings, and obtaining leads from prominent researchers and authors in this area (Wohlin, 2014). Once we completed the literature search, we comprehensively summarised the literature set in a spreadsheet, with sample entries shown in Tables 2 and 3. We then read the papers to ascertain if they all actually included surgical tool detection in some form or the other. For example, some of the studies on surgical workflow also included a surgical tool detection component since it has been reported that combining instrument signals with visual features leads to better segmentation, and faster and more accurate detection (Dergachyova et al., 2016). We discarded papers that did not discuss surgical tools or which used external markers for tool detection or tracking. The resultant collection of 161 papers are surveyed in this review (Fig. 2).

Dataset Review
Medical image analysis challenges have resulted in many new and innovative approaches to surgical instrument recognition. These challenges are designed to provide a platform for the development of cutting edge machine learning solutions in medical imaging, and research in these challenges has addressed instrument segmentation, detection and localisation, tracking and pose estimation, velocity and instrument state. Al Hajj et al. (2019) highlight the fact that more than twenty annual challenges were hosted, and the CATARACTS, EndoVis and M2CAI challenges specifically addressed the issue of instrument detection. In the medical image challenges, generally a specific task is defined, a dataset is provided, evaluation procedures are defined, algo-

Challenge Datasets
ROBUST-MIS 2019, a part of the EndoVis Challenge series, was based on surgical procedures from three types of surgery. The videos were from 30 minimally invasive surgical procedures: 10 rectal resection procedures, 10 proctocolectomy procedures and 10 sigmoid resection procedures. A labelling mask and instrument labels were manually created for the 10,040 extracted endoscopic video frames (Ross et al., 2019). This dataset was based on the Heidelberg colorectal data set (Maier-Hein et al., 2021). The Endoscopic Vision 2018 Robotic Scene Segmentation Dataset provided images that were based on actual surgical procedures and included considerable variability in backgrounds, instrument movements, angles, and scales. The entire challenge dataset was made up of 19 sequences of porcine endoscope images and the objective was to perform semantic segmentation of surgical images into a set of medical device classes and a set of anatomical classes (Allan et al., 2020). The EndoVis 2017 Robotic Instrument Dataset was made up of 10 sequences of abdominal porcine procedures, which presented seven different robotic surgical instruments (Table 4). The relatively small size of the dataset was an issue, since it was only made up of 3000 frames in total, out of which 1800 frames were selected as training data. The dataset supported three different segmentation tasks: binary segmentation, parts of instruments (e.g., shaft, wrist, claspers and ultrasound probes) and type segmentation (e.g., needle driver, forceps, scissors, sealer and others). The EndoVis 2015 instrument segmentation and tracking dataset provided data for rigid and articulated robotic instruments in laparoscopic surgery. For rigid instruments, 2D in-vivo images from four laparoscopic colorectal surgeries were provided for segmentation and in-vivo video sequences of four laparoscopic colorectal surgeries were provided for tracking. For articulated instruments, four 45second 2D images sequences of at least one large Needle Driver instrument in an ex-vivo setup were provided. Relevant annotations and additional test data were also provided. The Challenge on Automatic Tool Annotation for Cataract Surgery (CATARACTS) Dataset consisted of 50 videos of phacoemulsification cataract surgeries. Cataract surgery is the most common of the surgical procedures, and ophthalmologists use a wider range of tools than surgeons doing robotic or laparoscopic surgeries; consequently this dataset provided a large set of tools. There are more than nine hours of videos with an average duration of almost eleven minutes per surgery. A total of twenty one surgical tools are present in the videos (Table 4); a tool was only considered to be in use when in contact with the eyeball. In any particular frame, up to three tools can be visible at a time. However, this occurs in only 4% of the frames; 45% of the frames show no tools at all, 38% show one tool and 17% show two tools (Al Hajj et al., 2019) (Fig. 3).
The Cholec80 dataset contains 80 videos of cholecystectomy surgeries, and seven tools or instruments are present in the dataset (Table 4). Some tools-such as the grasper and hook-feature in many frames while other tools-such as the scissors and irrigators-are less used and appear with much lower frequency in the videos / frames (Twinanda et al., 2017). The m2cai16-tool dataset is a subset of the Cholec80  Dataset and it consists of fifteen cholecystectomy videos with binary annotations of the seven tools present (Fig. 4). Details of the surgical tool datasets used in the challenges is presented in Table 5. In addition to the metadata provided about each dataset, additional metadata charac-

Other Surgical Tool Datasets
In addition to the challenge datasets described above, many other surgical tool datasets have been developed and present these datasets in Table 6. Again, the blank cells in these tables serves to highlight the shortfall in metadata and details about these datasets. The ATLAS Dione dataset provided video data of ten subjects performing six different surgical tasks. The dataset was described as being challenging as it had camera movement and zoom, free movement of surgeons, a wide range of expertise levels, background objects with high deformation, and annotations that included tools with occlusions, change in pose and articulation or with partially visibility (Sarikaya et al., 2017). The Retinal Microsurgery (RMIT) dataset consisted of 18 in-vivo sequences of retinal procedures; for each sequence, four joints (Tip1, Tip2, Shaft and End Joint) of the retinal instrument were annotated. The RMIT was a single-instrument dataset-specified only as a Retinal Instrument. The dataset was further classified into four instrument-dependent subsets. There were three annotated tool joints and two semantic classes (tool and background). Lapgyn4 Dataset is a four-part gynaecological laparoscopy dataset comprising collections of images depicting general  5).
There were other datasets that we evaluated but did not include since they did not provide sufficient focus or coverage of surgical tools. These included DAISI: Database for AI Surgical Instruction (Rojas et al., 2020), the MISAW dataset used for the MIcro-Surgical Anastomose Workflow recognition on training sessions challenge (Huaulme et al., 2021), the Bypass40 dataset of laparoscopic gastric bypass procedures (Ramesh et al., 2021b), and the EAD2020 dataset   (Fig. 6). Liu et al. (2020a) highlighted the inconsistency in the terminology used in the research, and stated that terms are often differently defined and applied. Some of the terms which were used include detection, presence, localization, recognition, classification, identification, labelling and annotation. The taxonomy of terms used in the literature reviewed is presented in Fig. 7, it is clear that definitions and terminology varies considerably and there is no uniformity in the application or understanding of these terms. When computer vision tasks are considered, multiple problems have been addressed in the literature. Guo et al. (2016) discussed image classification, object detection, image retrieval, semantic segmentation, and human pose estimation as the key computer vision tasks. Chai et al. (2021) similarly listed the main applications as object detection or recognition, visual tracking, semantic segmentation, and image restoration, with image classification providing the basic backbone of each application. Voulodimos et al. (2018) evaluated object detection, face recognition, action and activity recognition, and human pose estimation in their survey of key tasks in computer vision. Al Hajj et al. (2019) state that these tasks can be categorized according the precision of the desired outputs, with the finest or more precise level of surgical tool-based tasks at the tool segmentation level. The next level of precision in tasks is tool localisation, and this often leads to either tool tracking or pose estimation. The coarsest task is tool presence detection or determining which tools are present in each frame of a surgical video. While we considered all these approaches, in actual practice a pipeline using all these types of algorithms would follow a logical flow of tool presence detection, tool localisation, tool tracking, tool segmentation and tool pose estimation. We therefore used this logical flow approach to structure our analysis of the research.

Tool Presence Detection Research
In work using the CATARACTS dataset, Roychowdhury et al. (2017) fine-tuned Inception-v4, ResNet-50 and two NASNet-A instances. In their solution, they relied on Markov Random Field (MRF) for modelling long sequences of approximately 20,000 frames. Sahu et al. (2017a) trained ResNet-50 initialised with ImageNet weights on this dataset. Prellberg and Kramer (2018) used the CATARACTS dataset to explore different ways to use ResNet-50, and reported that fine-tuning ResNet achieved consistently better results than using ResNet as a fixed feature extractor in combination with a custom classifier. Al  Twinanda et al. (2017) developed and used the Cholec80 dataset to test EndoNet, an architecture based on AlexNet, for tool detection. Sahu et al. (2017b) fine-tuned AlexNet on the m2cai16-tool dataset; using an approach similar to EndoNet. The Cholec80 dataset was used by Alshirbaji et al. (2018) to fine tune AlexNet for surgical tool classification. Mondal et al. (2019) used Cholec80 to train a multi-task learning framework based on ResNet50 trained on the ImageNet Dataset. The features extracted from the fully connected layer of ResNet50 were used to train a multitask Bi-LSTM. The final classification result was generated through combining the score results produced by both the LSTM hidden layers. Alshirbaji et al. (2021a) tested VGG-16, ResNet-50, DenseNet-121 and EfficientNet-B0 for surgical tool presence classification. This was tested on the Cholec80 and Cholec20 datasets. Alshirbaji et al. (2020a) generated synthetic data and used it to augment the Cholec80 dataset. AlexNet was fine-tuned using cross-dataset validation to improve tool presence detection. Vardazaryan et al. (2018) used ResNet18 pre-trained on Ima-geNet data and further trained the network on a Cholec80 sub-set of five videos annotated with image-level instru- Using the m2cai16-tool dataset, Raju et al. (2016) finetuned GoogleNet and VGG16 and used ten trained models (with 5-fold cross validation for both VGGNet and GoogleNet) in an ensembling process to obtain their final results. Zia et al. (2016) fine-tuned AlexNet, VGG-16 and Inception-v3 and presented a comparison of these different deep network architectures for surgical tool detection. Namazi et al. (2019) developed LapTool-Net, which was a contextual detector for surgical tools based on recurrent convolutional neural networks. The method exploited correlations among usage of tools in the m2cai16-tool dataset, as well as the context of the tools' usage for different tasks. Choi et al. (2017) proposed a real-time detection model for surgical instruments during laparoscopic surgery by using a CNN based on YOLO pre-trained on ImageNet. This was trained on the m2cai16-tool dataset. Hu et al. (2017) developed an attention-guided network (AGNet) and successfully tested it on the m2cai16-tool dataset. The method first extracted regions in images with high probability of containing surgical tools by a deep neural network (the global prediction network) and then analysed these regions via another deep neural network (the local prediction network) which provided a prediction for each tool. Lin et al. (2019) addressed surgical tool presence detection with the m2cai16-tool dataset as a multi-label classification problem. The authors relied on a pre-trained DenseNet201 with a classification layer whose output corresponds to the confidences of the presence of the seven tools in the image. Mishra et al. (2017) proposed a framework to detect tool presence in laparoscopy videos which consisted of a CNN based on ResNet50 for extracting visual features, and a Long Short-Term Memory network to encode temporal information. This was tested on the m2cai16-tool dataset. Leibetseder et al. (2018) used GoogLeNet (Szegedy et al., 2015) to classify images in the LapGyn4 dataset. Kletz et al. (2019a) used the Lapgyn4 Dataset for the task of binary classification to recognise video frames as either instrument or non-instrument image, and trained GoogLeNet for instrument classification. Murillo et al. (2018) developed a tree-structured convolutional neural network for the classification of 10 open surgery instruments. Eight separate CNNs were trained on ten surgical instruments, and four CNNs on five instruments. Murillo et al. (2017) used 5 open surgery tools for testing the performance of CNNs and Haar Classifiers (Viola & Jones, 2001) for surgical instrumentation classification. A tree based tool classifier was designed using four CNNs for presence detection of the five surgical instruments. Kurmann et al. (2017) presented a U-Net based surgical instrument detector which estimated instrument joint positions and instrument presence using a cross-entropy loss function. This was evaluated on a retinal and EndoVis 2015 datasets. Qiu et al. (2019) used the m2cai16-tool dataset and built a new dataset called the STT dataset with sequential frame annotations using bounding boxes. The authors then developed RT-MDNet, a real-time multi-domain convolutional neural network with three convolutional layers, a Region of Interest Alignment (RoIAlign) layer and three fully connected layers, and tested it on the STT Dataset. Hou et al. (2022) introduced an attention-based deep neural network-SKA-ResNet-composed of a feature extractor with a selective kernel attention module and a multi-scale regularizer to exploit the relationships between feature maps. Their SKA-ResNet was tested on a new surgical instrument dataset called SID19 for the classification of surgical tools.  Banerjee et al. (2019) used the CATARACTS dataset for a multi-label multi-class classification task, and developed a framework for localization and detection of tools. A tool counter was implemented using ResNet-18. Using activation maps, three smaller regions of interest were used to train a new CNN which predicted the tool type among the given 22 classes. Three baseline models were trained for the task-AlexNet, VGGNet and ResNet-18/50/152. Xue et al. (2022) proposed a pseudo supervised surgical tool detection (PSTD) framework, which used pseudo bounding box generation, box regressor, weighted mean boxes fusion and a classifier with bi-directional channel adaption for surgical tool detection. This weakly supervised surgical tool detection (WSTD) approach was successfully tested on the Cholec80 dataset using image-level tool category labels. Alshirbaji et al. (2021b) evaluated the generalisation ability of a VGG-16 model on images from different datasets for surgical tool detection. The datasets used were Cholec80 and a Gyna05 dataset which consisted of 5 videos of gynaecologic procedures, and target tools were the four surgical tools which were present in both datasets. Nwoye et al. (2021a) developed the CholecTriplet2021: the endoscopic vision challenge for the recognition of surgical action triplets in laparoscopic videos. The focus was on fine-grained surgical activity recognition, modelled as a triplet-instrument, verb, target. This was defined in terms of surgical activities as triplets of the actual instrument that was used, the actions performed, and the target anatomy for each surgery, and was provided as part of the EndoVis2021 grandchallenge. Nwoye et al. (2021b) developed a model which recognized triplets from these surgical videos by leveraging attention at two different levels-a Class Activation Guided Attention Mechanism (CAGAM) and a Multi-Head of Mixed Attention(MHMA). This method used cross and self attentions to capture relationships between the triplets. Nwoye et al. (2020) used class activation modules which used the instrument activation maps to guide the verb and target recognition. They used a dataset based on Cholec80 annoted with 135K action triplets-termed the CholecT40 dataset-and developed a multitask learning (MTL) network with three branches for the instrument, verb and target recognition. Liu et al. (2020b) used anchor-free convolutional neural network, based on a compact stacked hourglass network, for surgical tool detection. This was tested on the ATLAS Dione and Endovis Challenge datasets, and compared to results using Faster RCNN, Yolov3 (Darknet-53) and CenterNet (Hourglass-104). In surgical tool detection work associated with the ATLAS Dione dataset, Sarikaya et al. (2017) developed a framework with a Region Proposal Network (RPN) and a multimodal two stream convolutional network for object detection and localization, based on image and temporal motion cues. Fast R-CNN (Girshick, 2015) was used for the object detection task, and the region proposal boxes of RPN with the convolutional features were used as input for the detection network streams on both modalities. Using the EndoVis Challenge dataset and the ATLAS Dione dataset, Zhao et al. (2019a) adopted a frame-by-frame detection method using a cascading convolutional neural network (CNN) which consisted of two different CNNs for real-time multi-tool detection. The method was tested-along with Faster R-CNN (Ren et al., 2017), Yolov3 (Redmon et al., 2016), and RetinaNet (Lin et al., 2017)-on the two datasets. Liu et al. (2020c) proposed an anchor-free convolutional neural network (CNN) architecture using a compact stacked hourglass (Newell et al., 2016) Bouget et al. (2015) used the NeuroSurgicalTools dataset and developed a two step approach for surgical tool detection, where the first stage of the approach performed pixel-wise semantic labelling while the second stage matched global shapes. Leppanen et al. (2018) pioneered work for surgical instrument detection under high microscope magnification using CNNs in micro-neurosurgical videos. Two CNNs were trained-one for instrument detection and instrument tip location detection by classifying small parts of the frame at a time, and the second to detect whether the instrument is present in the frame using the full frame image. Law et al. (2017) trained a stacked hourglass network to detect the key-points of the robotic instruments in vesico-urethral anastomosis surgery videos using crowd-sourced annotations. They also trained a support vector machine (SVM) to classify the skill of a surgeon using the tracking results. Nakawala et al. (2019) used their Nephrec9 dataset to test a "Deep-Onto" network for surgical workflow and context recognition, including instruments. The network was an ensemble of deep learning models (Inception-V3 pre-trained on ImageNet) with knowledge management tools, ontology and production rules, including usage of instruments. This combined use of deep learning, knowledge representation and reasoning techniques was found to be effective for automatic surgical workflow analysis on robot-assisted urological surgery. Hossain et al. (2018) relied on CNNs for real-time surgical tools recognition in Total Knee Arthroplasty (TKA), and exploited region based convolutional neural networks to perform real time tool detection. The method was based on Faster R-CNN with VGG-16 as base network, and RGB image convolutional features were used to train a Region Proposal Network (RPN) that generated object proposals, the output was the coordinates of bounding boxes around the deployed surgical tools. Yamazaki et al. (2020) created a dataset from 52 laparoscopic gastrectomy videos, and used this to test Yolov3 for surgical instrument detection. Bar et al. (2020) used an approach based on inflating ResNet-50 into a 3D ConvNet model (I3D) for surgical phase classification. This was termed the short-term model, and the long-term model was a Long Short-Term Memory (LSTM) network. The approach used surgical tool presence as cues for each phase, and was tested on their laparoscopic cholecystectomy dataset.
Yang et al. (2019) relied on a Pyramid-UNet to localize a cardiac intervention instrument (RF-ablation catheter or guidewire) in a 3D ultrasound image for cardiac electrophysiology (EP) and transcatheter aortic valve implantation (TAVI) procedures. This was tested on their dataset of cardiac ultrasound images from porcine hearts. Colleoni et al. (2019) proposed a 3D FCNN architecture for surgical-instrument joint and joint-connection detection, using spatio-temporal features for robotic tool detection and articulation estimation. This was trained and tested on the EndoVis 2015 and the UCL dVRK datasets. Jin et al. (2018) extended the m2cai16tool dataset by providing labels for 2532 of the frames with the coordinates of spatial bounding boxes around the tools, and made a new m2cai16-tool-locations dataset available. Their approach for instrument localization was based on Faster R-CNN. In work that utilised the m2cai16-toollocations and m2cai16-tool-datasets, Jo et al. (2019) applied two algorithms-YOLO9000 (Redmon & Farhadi, 2017) and missing tool detection-to perform detection of surgical instruments in real time. Tang et al. (2022) leveraged multimodal imaging and deeplearning to dynamically detect surgical instrument positions in ophthalmic surgical maneuvers. In their system, they combined spectrally encoded reflectometry (SER) and crosssectional OCT imaging for automated instrument-tracking, and tested it on 4730 manually-labelled SER images of a 25-gauge internal limiting membrane ( (2019a) collected three phantom frame-sequence datasets using tracked surgical tools over an anatomical phantom. These datasets were used to test U-Net, TernausNet-11 with a pre-trained VGG-11 network, LinkNet-34 and LinkNet-152 for the semantic labelling, binary segmentation and real-time tracking of surgical tools without any human intervention.

Tool Tracking Research
Using a subset of the m2cai-2016 dataset, Zhang and Gao (2020) developed a surgical instrument tracking framework based on object extraction via deep learning, where a segmentation model extracted the end-effector and shaft of the surgical instrument in real time. The model was based on LinkNet with ResNet-18, pre-trained on ImageNet. Chen et al. (2017b) proposed a visual tracking method for surgical tool tracking based on a CNN with line segment detector (LSD) for the detection part and a spatio-temporal context (STC) learning algorithm for the tracking part. They successfully tested this system on three laparoscopic surgical datasets-a simulation dataset, a real in-vivo dataset and a standard dataset. Zhao et al. (2017) considered a surgical instrument as consisting of two parts: an end-effector and a shaft. Edge-points and line features were used for the shaft detection and a CNN based on AlexNet (Krizhevsky et al., 2012) was used to track and detect the end-effector. Hiasa et al. (2016) proposed and evaluated a method for segmentation of surgical instruments from RGB-D Endo-scopic Images using CNNs. The method used RGB and depth images from stereo endoscope images, and the output was a likelihood image, where white pixels indicated a high probability of instruments and black pixels indicated high probability of background. Segmentation was seen as a critical task for 3D surgical tool tracking and reconstruction. Zhao et al. (2019c) used two CNNs and six datasets to develop a coarse to fine method for surgical tool tracking. The first CNN, based on AlexNet, classified 10 surgical tool classes, and the second or fine CNN was a regression network for tracking of the tool tip area. This was tested on six different datasets-the first five were in-house surgical videos and the sixth was the Endo-Vis 2015 challenge dataset. Their method was compared with four other methods-Fast R-CNN with filter tracking in convolutional features using VGGNet, data-driven visual tracking, tracking with an active testing filter, and tracking with online multiple instance learning. Zhao et al. (2019b) developed an automatic real-time method for two-dimensional tool detection and tracking based on a spatial transformer network (STN) and spatiotemporal context (STC), and tested this on eight video datasets from in-house surgical videos. The authors tested their method and four other solutions-correlation filter tracking with convolutional features using VGGNet, datadriven visual tracking, tracking with an active testing filter and tracking with online multiple instance learning-on these datasets. Lu et al. (2020) tested a two deep neural networks framework for surgical tool tracking on the Surgical Perception (SuPer) and Hamlyn Centre Video Datasets. Using these datasets and a two CNN pipeline, a Pyramid Stereo Matching Network (Chang & Chen, 2018) was used to find and match features for stereo reconstruction, and DeepLabCut (Mathis et al., 2018) was used to detect point features for surgical tool tracking.

Tool Segmentation Research
For tool segmentation work, Luengo et al. (2021) added pixel-wise semantic annotations for anatomy and also surgical tools for 4670 images from 25 videos of the CATARACTS training set. This CATARACTS Semantic Segmentation dataset was used for the EndoVis 2020 challenge. Chen et al. (2021) developed a method that was based on exploiting cross-consistency in microscopic image segmentation, and used the consistency between the main decoder and auxiliary decoder to leverage unlabeled images. This was used to improve the Deeplabv3 plus network and was tested on the CATARACTS-Semantic-Segmentation 2020 data set. Zisimopoulos et al. (2017) used a FCN-VGG network that was trained to perform supervised semantic segmentation in 14 classes that represented the different tools present in their simulated cataract dataset. This dataset was used to train CNN models and then transfer learning techniques were used for training on the CATARACTS Dataset. Fox et al. (2020) used the CaDIS and the Cataract-101 dataset with Mask R-CNN to localize and segment surgical tools in ophthalmic cataract surgery. They compared four backbone networks (Inceptionv2, Inception-ResNetv2, ResNet50, and ResNet101-all with pre-trained COCO  weights-and different data augmentation strategies for multi-class instance segmentation of surgical tools. Grammatikopoulou et al. (2019) developed the CaDIS dataset for semantic segmentation in cataract surgery, based on the CATARACTS dataset. Pissas et al. (2021) highlighted that the main issue in using the CaDIS dataset was the extreme class imbalance in the granular semantic segmentation labels, and they addressed this challenge with two data oversampling strategies. They demonstrated that the choice of the loss function and data sampling strategy were paramount in training their ResNet based encoder-decoder networks.  Ceron et al. (2021) introduced a YOLACT architecture for real-time instance segmentation of surgical instruments, and tested its accuracy on the ROBUST-MIS dataset. They used criss-cross attention modules (CCAMs) with a ResNet-101 backbone to develop three models-CCAM-Backbone, CCAM-FPN and CCAM-Full-plus a baseline YOLACT++ model. Isensee and Maier-Hein (2020) relied on a 2D U-Net architecture that used residual blocks in the encoder and generated segmentation maps at several resolutions in the convolutional based decoder architecture. This method achieved a mean Dice score of 87.41 (94.35) on the ROBUST-MIS dataset. Sahu et al. (2021) used a teacher-student learning approach that learned from annotated simulation data and unlabeled real data. They redesigned their Endo-Sim2Real framework based on a teacher-student approach, and used a TerNaus11 as the backbone segmentation model. They tested this on a simulated dataset as well as on the Robust-MIS, EndoVis 2015 and Cholec80 datasets. Allan et al. (2020) reported segmentation results using the EndoVis18 dataset. The solutions included the use of the ResNeXt-101 architecture with Squeeze-Excitation blocks; U-Net architecture with a VGG 19 encoder; a global convolutional network (GCN) with ResNet 152 backbone; DeepLab V3+ using multi-scale feature extraction with Xception and atrous convolutions; WideResnet38 encoder and activated batch norm (ABN) with DeepLab V3 as decoder; two ResNet encoder blocks and a stacked convolutional decoder network with a sum-skip connection; 3 U-Net models with final prediction as an ensemble; a 77 layer fully convolutional dense network architecture; DeepLab V3+ and ResNet-50 pre-trained on ImageNet; a U-Net with a ResNet-101 backbone; and a Pix2Pix model for the segmentation with a U-Net as the generator. Most of the architectures were pre-trained on ImageNet. Gonzalez et al. (2020) extended the EndoVis 2018 dataset for fine-grained instrument segmentation by manually annotating each instrument in the dataset, and used this dataset to successfully test their ISINet model which was based on Mask R-CNN. Shvets et al. (2018) experimented with the U-Net, Ter-nausNet and LinkNet encoder-decoder architectures on the EndoVis 2017 dataset. TernausNet was shown to outperform the other architectures in all three tasks of binary, part-based and type-based segmentation. Hasan and Linte (2019) used U-Net but modified it to U-NetPlus model by introducing both VGG11 and VGG16 as an encoder with batch-normalized pre-trained weights and nearest-neighbour interpolation as the replacement of the transposed convolution in the decoder layer. This was tested on the EndoVis 2017 dataset. Mohammed et al. (2019) proposed a multi encoder and single decoder convolutional neural network, which they termed StreoScenNet. The architecture consisted of two ResNet50 encoder blocks, pre-trained on ImageNet, and a stacked convolutional decoder network connected with a sum-skip connection. The input to the encoder was a set of left and right frames, and the output of the decoder was a mask for the instrument, part and binary segmentation tasks. This was tested on the EndoVis 2017 dataset. Zhang et al. (2021b) proposed a GAN-based method for unpaired imageto-image translation (I2I), and used it for surgical tool image segmentation and repair. They tested this on three endoscopic surgery datasets and on the EndoVis17 dataset. Kong et al. (2021) optimised Mask R-CNN model with anchor optimization and improved Region Proposal Network for surgical instrument segmentation. They evaluated their architecture on the EndoVis17 and an in-house hysterectomy dataset. Kurmann et al. (2021) proposed a encoder-decoder network for segmentation and classification of surgical instruments in endoscopic images. Their "segment first, classify last" approach used a shared encoder, two decoders for instance segmentation, and a classifier for instance classification, and it provided good results on the EndoVis 2017 dataset. Ni et al. (2019) introduced a Refined Attention Segmentation Network (RASNet)-based on ResNet-50 pre-trained on ImageNet-to simultaneously segment and classify surgical instruments. An Attention Fusion Module (AFM) was used to fuse multi-level features by utilizing the global context of high-level features as guidance information, and this was tested on EndoVis 2017. Islam et al. (2019) developed a light-weight cascaded convolutional neural network to segment surgical instruments from the EndoVis 2017 data. The authors developed a Multi-resolution Feature Fusion (MFF) block to fuse feature maps from their auxiliary and main branches, and combined auxiliary loss and adversarial loss to regularize the segmentation model. A spatial pyramid pooling unit was used to aggregate rich contextual information in their intermediate stage. Islam et al. (2021) proposed a Spatio-Temporal Multi-Task Learning (ST-MTL) model with a shared encoder and spatio-temporal decoders for real-time surgical instrument segmentation and tested it on EndoVis 2017. Comparative tests were also conducted on other models using identical pre-processing and augmentation techniques. Lee et al. (2019b) presented a "Two-phase Deep learning Segmentation for Laparoscopic Images" (TDSLI) model and tested it on the EndoVis 2017 dataset and an additional dataset of four retrospectively collected laparoscopic image sequences in different animal surgeries. The LinkNet-34 network was used in a convolutional encoder-decoder architecture, with a pre-trained ResNet-34 network used for the encoder. Jha et al. (2021a) released the "Kvasir-Instrument" dataset with annotated bounding box and segmentation masks of GI diagnostic and surgical tools, and tested it using the U-Net and DoubleUNet architectures for semantic segmentation. Andersen et al. (2021) reported the success of Mobile-U-Net for the segmentation of surgical tools and suture needles, and tested it on a laboratory dataset and JIGSAWS (Gao et al., 2014) dataset. Choi et al. (2021) used the YOLOv4 and YOLACT-based models for real-time object detection and semantic segmentation of six surgical tools in a mastoidectomy surgery dataset. Zadeh et al. (2020) used a gynaecological dataset to train Mask R-CNN, which was then tested on laparoscopic images from 2 additional surgeries not included in the training set. Qin et al. (2020) Kletz et al. (2019b) used a ResNet50 architecture as a backbone network with a feature pyramid network (FPN) for instance segmentation task using images of gynaecological surgeries. They also fine-tuned a Mask R-CNN  model for seven instrument classes (including "BG" or Background) using a pre-trained model on the COCO dataset. VGG, PSP (Zhao et al., 2016), UPerNet (Xiao et al., 2018) and DeepLab (Chen et al., 2016) were trained and evaluated for anatomical understanding, instrument identification and tracking, and understanding of interactions between surgical instruments and anatomical landmarks. Sahu et al. (2020) used two datasets-Cholec80 and EndoVis 2015-to test their Endo-Sim2Real method for instrument segmentation. TerNaus11 was used as the DNN model for the instrument segmentation task. Kanakatte et al. (2020) proposed a pixel-wise instance segmentation algorithm for the segmentation and localisation of surgical tool using a spatio-temporal deep network, and tested it on Cholec80. Their model used ResNet pre-trained on ImageNet database and Inflated Inception 3D (I3D) pre-trained on the ImageNet and Kinetics datasets (Kay et al., 2017) to capture spatio-temporal features. They also implemented and tested U-Net and Mask R-CNN on their annotated Cholec80 dataset. Laina et al. (2017) modelled the tool segmentation and pose estimation problem as a heatmap regression where every pixel represented a confidence proportional to its proximity to the correct landmark location. For encoding, ResNet-50 pre-trained on ImageNet was used and three different CNN variants were defined for the decoding task. The model was tested on the RMIT and EndoVis 2015 datasets. Du et al. (2018) added detailed annotations to existing labels for the RMIT and EndoVis 2015 datasets, and tested a framework with a fully convolutional detection-regression network for articulated multi-instrument 2-D pose estimation. Kayhan et al. (2019) proposed a lightweight deep attention based network architecture and evaluated three SSL algorithms for a deep attention based semi-supervised 2D-pose estimation method for urgical instruments: mean teacher, virtual adversarial training and pseudo-labelling. Analysis was conducted n the RMIT and EndoVis 2015 datasets. A modified U-Net architecture (DAU-Net) that made use of attention echanisms was used to find each tool joint location via a heatmap output channel. Kugler et al. (2020a) introduced three datasets: two synthetic Digitally Rendered Radiograph (DRR) Datasets (the first with a screw and the second with two surgical instruments), and a real X-ray Dataset (with manually labelled screws). They used this for a three step approach for surgical pose estimation including the application of a convolutional neural network based on a VGG architecture for information extraction, and then pose reconstruction from pseudo-landmarks. Kugler et al. (2020b) used two of these datasets to test an automatic framework (AutoSNAP) for the discovery of neural network architectures for instrument pose estimation, leading to the development of an improved architecture (SNAPNet). Hasan et al. (2021) developed a CNN they called ART-Net, for Augmented Reality Tool Network, and combined it with an algebraic geometry approach for generic tool detection, segmentation, and 3D pose estimation. While the CNN ART-Net was used for surgical tool detection and segmentation, geometric primitives were also extracted to compute the 3D pose with algebraic geometry. Gessert et al. (2018) addressed surgical tool pose estimation from optical coherence tomography (OCT) volume data with a deep learning-based tracking framework called Inception3D. The 3D CNN architecture was used to learn accurate regression between volumetric images and object poses, and was then used to estimate object pose from new volumetric images.

Open Research Questions
We address our research questions by presenting a comprehensive review of surgical tool datasets. A knowledge hierarchy of machine learning research was then developed using these datasets. However, while robustness or the reliable performance of methods on challenging images has been addressed in the work, there are important questions and research gaps that need to be addressed. These issues are discussed in this section.

Data Modalities
As we have found in our survey, RGB images or video are the predominant data modalities in the datasets. This is a well understood modality, and it is easy to deploy cameras to capture entire room images, high level views of the procedures, specific images of body parts, or even for internal imaging through endoscopes . However, there are many more medical modalities that can be explored for creation of rich and representative datasets. A limited amount of work using other images modalities is reported, and this includes radiograph and X-Ray (Kugler et al., 2020a), optical coherence tomography (OCT) (Gessert et al., 2018), RGB-D depth (Hiasa et al., 2016), and 3D ultrasound images (Yang et al., 2019). Multi-modal datasets could potentially be valuable-for example, in their review of surgical activity recognition research, van Amsterdam et al. (2021) reported that multi-modal data integration demonstrated promising results on small surgical datasets. While image modalities tend to be specific to surgical areas, there are some modalities that could foster innovative work in the surgery domain-for example, the use of IR images to supplement standard RGB images could address issues with illumination and reflection, and could lead to more accurate models being developed. Similarly, depth images could assist in addressing surgical tool counting problems and for segmenting tools from complex and crowded backgrounds.

Dataset Volume, Variety and Quality
In a white paper on the first annual Conference on Machine Intelligence in Medical Imaging (C-MIMI), Kohli et al. (2017) discussed the impact on machine learning performance due to the unavailability of large and high-quality training data. The lack of data for medical image evaluation with machine learning is a key concern, to the extent that the term "data starved" was used to describe the state of current research in this area. Similarly, van Amsterdam et al. (2021) stated that the availability of large and diverse open-source datasets of annotated data was essential for the development and validation of robust solutions in the surgery domain. A further challenge in medical surgery domains is the great variety of surgeries and the rapid rate of change (i.e. new techniques and tools) which increases the chance that a medical dataset will become obsolete, a problem that is generally not present in traditional object detection domains.
In a workshop on Surgical Data Science (SDS), Maier-Hein et al. (2020) discussed the lack of success stories in surgery, and contrasted it to success with machine learning research in other medical areas, such as radiology, dermatology, gastroenterology and mental health. This lack of success was directly attributed to the lack of quality annotated data, representative of the surgery domain. Participants in the workshop cited the EndoVis, Cholec80 and JIGSAWS datasets as being useful for research but the small size and limited representation provided by the datasets-even in these major initiatives-was reported to be a core issue. It was stated that creating and providing access to larger, more-representative and fully annotated datasets would lead to improved outcomes and success stories in the application of machine learning to surgery. Bouget et al. (2017) reviewed the surgical tools used in different setups and for different procedures and found that two categories of surgical tools emerged: articulated instruments and rigid instruments. This survey also found two such categories into which most works fell-we categorise them as either laparoscopic instruments or open surgery tools. Table 7 indicates that the overwhelming majority of work in this area has focused on laparoscopic surgery, and open surgery has received considerably less attention. Even the work that has been accomplished in open surgery focuses on very few instruments; the majority of work detects less than 10 instruments and even the Cataracts dataset provides only 21 instruments (Table 1). There are tens of thousands of instruments in circulation in a hospital at any one time and we would also expect tools to change over time or new tools to be introduced due to new technology or innovations in surgical techniques. Clearly, therefore, larger datasets are required and it would be useful for the research community if more open surgical tool datasets are made available.
Ideally, a surgical tool dataset should have large data volume, expert annotations, reliable ground-truth, and reusability. An issue is the size of available datasets, the benchmark dataset-ImageNet-has 14 million categorized images in a hierarchical arrangement. By contrast, most medical image datasets are limited to hundreds of cases, and datasets with thousands of annotated images are very limited . A valuable initiative would be to create and curate a large surgical tool dataset of tens of thousands of tool images across surgical specialities with different modalities of image capture. Further, all the datasets surveyed in our paper have a flat structure. Given that fact that surgery is organised along specialities (Table 7), and each speciality has separate underlying categories, a hierarchical classification of surgical tools in the datasets provided for machine learning research has been shown to be extremely valuable (Rodrigues et al., 2022(Rodrigues et al., , 2021a.

Dataset Bias and Generalisation
A major problem highlighted by Barbu et al. (2019) is that most datasets are highly biased. The objects of interest were generally highly correlated with the image backgrounds and objects were presented in stereotypical orientations with limited occlusions and under standardised illumination conditions. These biases were problematic because training on these datasets did not transfer well to real world data where there were variable views, orientations, backgrounds and illumination (Barbu et al., 2019), and there is limited research that tests or addresses this problem. In our survey, we found that benchmark datasets capture very specific image types with similar backgrounds, modalities, controlled collection methods, identical contexts and annotations. A key concern expressed in the literature is about algorithms which are trained on a specific dataset, procedure, intervention or in specific institution being able to generalise to other datasets and procedures (Ross et al., 2019).
To ensure viewpoint invariant object detection, different angles, scales, background clutter, illumination, orientation, pose, occlusion and intra-class variations should be captured in the images. Generalisation can be estimated by conducting research across different datasets using the same model. For example, Sahu et al. (2020) tested the Endo-Sim2Real model for instrument segmentation across two datasets-Cholec80 and EndoVis 2015, Zhao et al. (2019a) tested their method on the EndoVis Challenge dataset and the ATLAS Dione dataset, and Kalavakonda et al. (2019) evaluated three different deep architectures-U-Net, VGG16 and MobileNetV2-on their NeuroID dataset and on the EndoVis 2017 dataset. Du et al. (2018) and Kayhan et al. (2019) developed machine learning solutions and tested them on the RMIT and EndoVis 2015 datasets. More research initiatives across datasets to evaluate issues such as how accuracy or performance changes from one dataset to another, or the dependence of performance on camera or image quality, is essential.
More research is also required across the fourteen surgical specialities as listed in Table 7, since the current research is limited in scope and scale and only addresses a few specialities, but to accomplish this, better surgical tool datasets need to be made available.

Issues with Annotations
Maier- Hein et al. (2014) highlighted the fact that the performance of deep learning classifiers are heavily dependent on the availability of relevant annotations, and point out that such annotations are difficult and expensive to obtain because they need medical expertise and experience. Since medical resources for this task are limited, available datasets for deep learning are typically small and unable to cover the required range of variance for training deep learning systems for medical applications. Orting et al. (2020) hypothesised that the high costs associated with annotations is a factor in the limited availability of large-scale, well-annotated datasets. They reviewed 57 papers that used crowd-sourcing for the analysis of medical images and for labelling large quantities of data. They reported that 42% of the papers they surveyed focused on classification, 39% on localisation or segmentation, 12% on both classification and segmentation, and a further 7% on other tasks-each task required specific annotations to be performed, with varying degrees of complexity and difficulty. Hein et al. (2018) state that deep learning based techniques for medical applications require huge amounts of accurate reference segmentation annotations, and completing manual annotations is extremely time consuming. The authors state that crowd-sourcing could result in accurate and costeffective annotations for radiology images, and showed that even non-experts were able to complete high quality image segmentation in the medical domain.   Nogueira-Rodriguez et al. (2020) reported that the publicly available datasets that could be used for object detection all annotated the object locations as binary masks. These masks were directly used for deep learning solutions but could also be converted to bounding boxes if required for specific training strategies. Annotation costs also vary across types of surgery-for example, annotation of surgical tools in cataract surgery needs to specify if the tool is actually in use or in contact with the eyeball, and this requires expert annotators to define (Al Hajj et al., 2019). This is expensive and tedious, but other surgery types only define the presence of the object in the frame, therefore needing simpler, cheaper annotations. In general terms and as Garcia-Peraza-Herrera et al. (2021) point out, manual annotation of pixel-level segmentation labels is difficult, expensive, tedious and timeconsuming, this has led to a shortfall in the availability of quality datasets for deep learning. Since there are no large datasets available for tasks such as deep learning based surgical instrument-background segmentation, advancement in this area has been significantly curtailed. Ward et al. (2021a) discussed the challenges in annotating spatial, temporal, and clinical elements of surgical videos, and in achieving consistency and reliability of annotations across the data. They also highlighted the requirement for achieving consensus in the development and use of surgical annotations. Meireles et al. (2021) studied current practices in surgical video annotation, and proposed recommendations for the annotation process. This is an on-going effort to create a general framework of recommendations to facilitate uniform annotations and to improve cross-institutional research efforts. Initial recommendations appear to call for increased detail in annotation-for example, to include hierarchical information of surgical tools, anatomy, and tissue types, as well as for patient-specific factors and intra-operative influencing factors in the annotations. Kohli et al. (2017) pointed out that there are no generally accepted standards for the creation and cataloguing of medical image datasets. As we demonstrate in Table 6, surgical tool dataset collection, curation and use is typically provided as a one-off solution, directly linked to a specific research project. The metadata provided with these datasets, if at all available, is all too often limited in description, incomplete and inconsistent. Specific domain and speciality expertise as well as knowledge of the context and institution is required to make sense of the data provided. In our Table 5, we provide metadata for the important publicly available machine learning datasets that address surgical tool tasks, more information would be useful and this is perhaps a starting point for future work to make datasets more understandable and useful (Kohli et al., 2017).

Metrics
There are an extremely wide range of metrics that have been used in the research (Fig. 8). Reinke et al. (2018) reported 14 different metric used by the MICCAI in 75 grand challenges held between 2007 and 2016. The range of metrics, variety of approaches and different reporting criteria made it difficult to directly compare results. For example, Zhang and Gao (2020) reported sensitivity, specificity, dice similarity coefficient (DSC) and model inference time (MIT) for their work on the m2cai2016 dataset, while other researchers reported the Mean Average Precision. Zia et al. (2016) tested AlexNet, VGG and Inception of the m2cai2016 dataset but pointed out that comparisons were not fair since the first two architecture were tested by removing one of the 10 videos, while the third architecture was tested by randomly selecting a percentage of the input data for testing and validation. A standard set of metrics, consistent and fixed splits of datasets into, for example, training, validation and testing, and standard metrics for evaluation would be useful for future research but it is difficult to make a hard recommendation since this is very task and context specific.

MLOps and Federated Learning
Given the mission critical nature of surgical tool management in a hospital, the deployment of deep learning systems in real time-or MLOps-needs to be addressed (Makinen et al., 2021). We have highlighted the tremendous progress that has been made in the application of deep learning models to surgical tool management in this survey, but the deployment, integration, adoption and testing of such systems in actual hospital conditions remains a significantly under-explored area due to the lack of data, the general messiness or poor usability of data, and the inaccessibility of data (Makinen et al., 2021). Making sure that consistently high-quality data is available for MLOps, while ensuring coverage of all data cases and creating data annotations that are consistent, is therefore a critical task (Ng, 2021).
Given the fact that the surgical tool datasets used for deep learning are generally small in size, private in nature and distributed across many institutions, federated learning may offer a way to overcome the size and accessibility barrier. With federated learning, local data can be used for local training, and this can then be aggregated with other locally trained models for deep learning (Zhang et al., 2021a). Rieke et al. (2020) highlighted the fact that health related data is difficult to obtain, sensitive in nature, strongly controlled by privacy and other regulations, is expensive to collect, curate and maintain, and therefore generally not available on the scale needed for training deep learning models. Whatever medical data is available tends to be very task-or disease-specific, and of limited utility given license restrictions. Demonstrating the practicality of this approach for biomedical research, Silva et al. (2019) developed a federated learning framework for the analysis of multi-centric, multi-database sub-cortical brain data. Table 8 summarises the open research questions and opportunities which we identified and detailed in previous sections of this paper.

HOSPI-Tools Dataset
The currently available datasets used for surgical tool recognition offer a limited range of instruments to work with, with a maximum of 21 instruments, but-as we have identified in our review-better datasets are required for research. To help in addressing these challenges, we created a new surgical tool dataset named HOSPITools-"Hierarchically Organised Surgical Procedure Instruments and Tools" (Rodrigues et , 2022, 2021a, b). We created an initial dataset of surgical instrument images: over forty thousand images of surgical tools were captured using under different lighting conditions and with different backgrounds. Meireles et al. (2021) point out that surgical instruments can present significant differences due to their function, and intended possible uses, as well as due to manufacturing variations. They therefore recommended hierarchical annotation at two levels-the general and the specific instrument type-so that research can address device-related complications or surgical issues stemming from any particular device, the outcome from specific instrument choices, and the use of instruments in different surgical procedures. Since instruments could be used for multiple purposes, the authors recommended that additional labels be added to instrument annotations. We instead built the hierarchical structure directly into our dataset and created a four level hierarchy which consisted of speciality (2 classes), pack (12 classes), set (35 classes) and tool (360 classes) levels. We believe that this approach can be valuable for deep learning research and this dataset was therefore designed to offer a large variety of tools, arranged hierarchically to reflect how surgical tools are organised in real-world conditions. We provide details of the HOSPI-Tools Dataset  Table 9, and examples of actual instrument sets and annotations of instruments in Figs. 9 and 10. Images captured included individual object images as well as cluttered, clustered and occluded objects. More images need to be taken by adjusting the DSLR camera position and pose-this would increase the realism and utility of the dataset. Instrument images were captured before and after use in surgery, it was not possible to take images of the tools in use during actual surgery. Our survey findings have highlighted the need to include more images with occlusions, illumination changes, and the presence of blood, tissue and smoke, to accurately capture complex surgery conditions. This is one step in the direction of addressing the issues that we have identified in this survey, but much more work needs to be accomplished. We will add other specialities as we develop this dataset, to reflect the complexities inherent in each of the surgical specialities and to address the open research issues and challenges.

Conclusions
We presented a comprehensive survey of datasets for surgical tool detection and related surgical data science and machine learning techniques and algorithms. We offered a high level perspective of current research in this area, analysed the taxonomy of approaches adopted by researchers using surgical tool datasets, and addressed key areas of research, such as the datasets used, evaluation metrics applied and deep learning techniques utilised. To ensure that we were rigorous and structured in our approach, we defined an a priori protocol for discovering and selecting the research that we reviewed. Adherence to this protocol prevented any mid-stream shifting of goals and inclusion criteria, and ensured that we presented a comprehensive and robust knowledge hierarchy.
Our survey shows that the application of machine learning to surgical tool detection, localisation, tracking, segmenting and pose estimation is a well explored research subject and many innovative techniques have been applied. However, we also identified and discussed the open research issues and challenges. To help address some of the gaps and shortfalls that we have identified, we make a contribution by creating a new Surgical Tool Dataset and we make this dataset publicly available to encourage more work in this direction.

Declarations
Conflict of interest The authors declare that they have no conflict of interest.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecomm ons.org/licenses/by/4.0/.