Introduction

Magnetic resonance imaging (MRI) is the first imaging choice for detecting and localizing prostate cancer [1, 2], based on the Prostate Imaging Reporting and Data System (PI-RADS) scoring system [3] and depending on zonal anatomy. Zonal segmentation of the prostate plays a crucial role for prostate cancer detection as the PI-RADS score differs depending on the areas studied, based on diffusion-weighted imaging (DWI) for peripheral zone lesions and T2-weighted (T2W) imaging for transitional zone lesions, but also for multiple clinical application such as reproducible prostate volume and Prostate Specific Antigen (PSA) density evaluation [4], MRI-ultrasound fusion biopsy, radiotherapy, or focal planning.

Zonal segmentation of the prostate is usually performed manually on T2W images by contouring the prostate in a slice-by-slice manner. It is extremely time-consuming, tedious, and prone to inter and intra-observer variability due to the subjective human interpretation of organ boundaries and large variability in prostate anatomy and gland intensity heterogeneity across patients [5]. There is a real need to develop automatic methods to accelerate the whole process and offer robust and accurate prostate segmentation.

Automatic zonal segmentation of the prostate is a challenging task for multiple reasons. Prostate gland is subject to large morphological variation, intra-prostatic heterogeneity, and poor contrast with adjacent tissues, making delineation of prostatic zonal contours laborious. Multi-institutional applicability can be difficult to evaluate as there is a wide technically induced variability in the image acquisition, as MRI signal intensity is not standardized and image characteristics are strongly influenced by acquisition protocol, field strength, scanner type, coil type, etc. [6].

Finally, the performances of an automated segmentation method depend in part on the database (heterogeneity of the data used, knowledge of possible selection biases), quality of ground truth (manual delineation of the prostate performed by human experts), training time and hardware requirements. First commonly used methods were based on machine learning methods, such as atlas-based registration models in which several reference images with corresponding labels are registered and deformed onto the target image [7, 8] or C-means clustering models [9, 10]. Most common methods described after 2017 are based on deep learning with convolutional neural networks (CNN) allowing automatic extraction of features and semantic image segmentation. Common architectures such as U-net [11], V-net or ResNet have been extensively used. Modification and fine tuning of existing models, by either combining multiple U-nets [12,13,14], adding attention modules such as squeeze and excitation [15], feature pyramid attention [16], adding blocks [17], transition layers or up-sampling strategies [18], allowed either improving accuracy of classical CNN or obtaining same accuracy with reduced memory and storage requirements.

The primary objective of this review was to provide a detailed analysis and comparison of applicability and efficiency of the published methods for automatic segmentation of prostate zonal anatomy by systematically outlining, analyzing, and categorizing the relevant publications in the field to date. We also aimed to identify methodological flaws and biases to demonstrate the need for a consensus on quality criteria for database constitution and prostate zonal segmentation methodology.

Materials and methods

This systematic review was conducted and reported in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses statement (PRISMA) [19]. The methods for performing this systematic review were registered on PROSPERO [20] database (registration number CD42021265371), and were agreed by all authors before the start of the review process to avoid bias. This study was exempt from ethical approval at our institution because the analysis involved only de-identified data.

Data sources and search

Medical literature published in the English language published until 30 June 2021 was searched in multiple databases (Medline, Science direct, Embase and Web of Science) using the following terms:

(prostatic OR prostate) AND (automated OR automatic) AND (segmentation OR segmented) AND (zone OR zonal) AND (\"magnetic resonance\" OR mri OR \"magnetic resonance\"OR mri OR mr) AND (\"artificial intelligence\" OR \"deep learning\" OR \"machine learning\") and all possible combinations.

No beginning date was applied.

Study selection

Full-text selection was independently performed by two radiologists, one experimented radiologist specialized in uroradiology and prostate imaging (S.M., 5 years in prostate imaging, with more than 1000 cases of prostate MRI per year) and one radiology fellow specialized in uroradiology and prostate imaging (C.W., 1 year in prostate imaging, with more than 1000 cases of prostate MRI per year). A third experimented professor of radiology specialized in prostate imaging (R.R.P., 15 years in prostate imaging, with more than 1000 cases of prostate MRI per year) intervened in case of disagreement. We summarized search strategy details for each database in Fig. 1.

Fig. 1
figure 1

Flow diagram based on Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) recommendations for systematic reviews

Full size image

We imported all articles retrieved into the reference manager Zotero and removed all duplicates. The same two radiologists (C.W., S.M.) then independently and manually screened titles and abstracts of the resultant database to ensure relevance. Articles that were obviously out of the scope of the research topic were excluded at this stage. Subsequently, all the remaining articles full texts were retrieved and read, applying inclusion and exclusion criteria (explained below) with conflicts resolved by consensus with the third reviewer. Reference lists of these relevant articles were also reviewed for possible papers missed in the primary search, and those papers were screened using the same initial inclusion and exclusion criteria.

Selection criteria

Inclusion criteria

Articles were included if they were original articles, used machine learning or deep learning algorithms and aimed to segment prostate human MRI images by zonal anatomy, using a fully automated method with manual segmentation as ground truth.

Exclusion criteria

Articles were excluded if they were commentaries, editorials, letters, case reports or abstracts. Were also excluded articles with semi-automated segmentation methods, no description of segmentation method, segmentation of the whole gland (WG), or prostate cancer without zonal anatomy, absence of similarity metrics or of evaluation against ground truth segmentations.

Data collection and extraction process

The qualifying papers were then reviewed, and various data of the studies were extracted and tabulated prior to analysis (Table 1).

Table 1 Data extraction
Full size table

Assessment of methodological quality

The two same radiologists (C.W., S.M.) independently assessed and extracted data from each of the included articles, using the Quality Assessment of Diagnostic Accuracy Studies tool-2 (QUADAS-2) framework [21] adjusted with topics from the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) [22] to evaluate the risk of bias and applicability for each selected study, with conflicts resolved by consensus with the third reviewer.

Extracted data were tabulated, synthesized, and evaluated for methodological flaws and applicability of the proposed techniques.

Results

After removing duplicates, 458 articles were remaining. Final consensus was reached yielding a total of 33 articles [6,7,8,9,10, 12,13,14,15,16,17,18, 23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43] (Figs. 1, 2).

Fig. 2
figure 2

Chronological distribution of the 33 reviewed articles. 1st model for prostate zonal anatomy segmentation was published in 2011. 1st convolutional neural network (CNN) was published in 2017

Full size image

Datasets

Training, validation, and test sets

All articles used retrospective datasets.

Wide heterogeneity in training, validation and test datasets was found (Table 2).

Table 2 Overview of types of databases used with training, validation and test sets distribution
Full size table

Performance testing of the algorithms can be done on same source than for the development or use different source of data, and based on either public data, private data or a combination of both. Public data were used in 15/33 articles for testing. Only 7 studies [6, 9, 14, 30, 33, 36, 37] used both private and public data for testing, allowing better generalizability of their algorithms. None of them used prospective data for validation and testing.

Most used public datasets were PROSTATEx [44], NCI-ISBI 2013 [45] and PROMISE12 [46] (Additional file 1: Table S1).

Eight authors applied cross-validation, using a subset of available dataset as training set, while the remaining data constituted the test set to evaluate the segmentation performance and accuracy. Nine reported using cross-validation for testing, averaging the results from the different rounds, hence adding bias.

Technique

We identified major technical differences in datasets regarding the number of vendors, field strength, type of coils, sequences, slice thickness, field of view (FOV) and input data used for automatic segmentation (Table 3). Less than half (14/33) studies used more than one type of vendors and 7/33 used both 1.5T and 3T MRI machines. More than 2/3 (24/33) used mono-modal input, mainly T2-weighted planes, in combination with apparent diffusion coefficient (ADC) map in one study [13] or with multiparametric and multi-incidence MR images in another [9]. The slice thickness of T2-weighted axial planes was consistent with the PI-RADS v2.1 recommendations in 13/33 studies (≤ 3 mm), which was not the case for the public data base PROSTATEx (3.6 mm). Only 7 studies provided sequence details (type of sequence, slice thickness, FOV) used for ground truth manual segmentation.

Table 3 Input MRI parameters (number of vendors, type of field, type of coil, input sequences)
Full size table

Zonal anatomy

We found 18 different types of very heterogeneous and unclear terminologies of zonal anatomy (Fig. 3, Additional file 1: Fig. S1). Out the 33 articles reviewed, less than 1/4 (8/33) [23, 25, 32, 34, 36, 37, 40, 43] provided precise terminology and segmentation protocol. Frequently the inappropriate term “central gland” (CG) was used, with ambiguous definition of central zone (CZ) and anterior fibro-muscular stroma (AFMS) alternatively included in peripheral zone (PZ) or transition zone (TZ), or mainly not described at all. Two studies mis-used the term “central zone” to refer to the “central gland” [27, 39].

Fig. 3
figure 3

Schematic of the four major types of protocol of zonal segmentation. Type A: articles for which “central gland” included CZ, TZ and AFMS. Type B: articles for which “central gland” included TZ and CZ. No details for AFMS. Type C: articles which did not provide details for AFMS, CZ or CG. CZ seemed to be mostly segmented PZ, while AFMS seemed to be mostly segmented with TZ, usually called “CG”. Type D: articles which did not provide details for AFMS or CZ. CZ and AFMS seemed to be mostly segmented with PZ. CZ central zone, TZ transition zone, AFMS anterior fibro-muscular stroma, PZ peripheral zone, CG central gland

Full size image

Ground truth

Manual delineation of the prostate gland performed by human experts was used to generate ground truth (Table 4).

Table 4 Type of ground truth segmentation
Full size table

Annotation tool

Twenty studies (61%) reported using manual contouring, while a third (11/33) reported using annotation tools. One team [31] specified that the radiologist did not delineate zones on all slices but relied on interpolation performed by their annotation tools. Two studies [32, 33] did not provide any information.

Qualifications of annotators

Most studies (27/33, 81%) reported a radiologist or a radiation oncologist as human expert. In 3 papers, no detail was provided on annotators qualification, although one [15] specified using an “expert” reader. Definition of an “expert” reader was mostly unclear with no specification of number of MRI they interpreted, for example [10, 15, 26, 31, 34, 39].

Number of readers

Number of readers and their experience are described in Table 4. Number of readers was not available in two studies. While 2/3 of teams (22/33) reported using more than one reader, with splitted, stratified or blinded reading approaches, 7 did not provide information on reading approach.

Intra and inter-rater variability

Inter-rater variability for annotations was rated in only 4 studies [7, 10, 23, 39]. Some studies used alternative techniques to approach better homogeneity of ground truth. In [13], the four radiologists met for a training session and together segmented two example patients to achieve a similar methodology for the rest of the dataset, using only experienced radiologists. In [6], the contours segmented by three radiologists were cross-checked and reviewed by two radiation oncologists, resulting in better homogeneity of ground truth. In [18], the initial prostate masks were drawn by two students who were trained in segmenting prostate zones.

Risk of bias and quality assessment

The detailed results are presented in Fig. 4 and Additional file 1: Table S2.

Fig. 4
figure 4

Stacked bar charts showing results of quality assessment for risk of bias and applicability of included studies. QUADAS-2 scores for methodologic study quality are expressed as the percentage of studies that met each criterion. For each quality domain, the proportion of included studies that were determined to have low, high, or unclear risk of bias and/or concerns regarding applicability is displayed in green, orange, and blue, respectively. QUADAS-2: Quality Assessment of Diagnostic Accuracy Studies 2

Full size image

Regarding patient selection, we considered a low risk of bias if there were clear data inclusion and exclusion criteria, inclusion of patients with and without PCa. Models were considered less applicable if datasets were composed of only one type of scanners or if no information was specified.

For reference standard, number of readers and type of reading for ground truth segmentation were reviewed.

Clear partitioning of the database (into training, validation, and test sets) was needed to waive risk of bias for flow and timing. Some articles used cross-validation methods without keeping a clear independent test dataset [6,7,8, 15, 25, 26, 30, 33, 36].

Overall, all 33 included studies were judged to have a low risk of bias in the domain “index test” and 22 of 33 (67%) of the studies were judged to have a low risk of bias considering “flow and timing”. However, only 1/4 of the studies (8/33) were judged to have a low risk of bias in the domain “patient selection”, 1/3 (10/33) in the domain “reference standard”. Only 2 articles were judged to have a low risk of bias in all four domains.

AI methodology

Before 2017, authors mostly used machine learning-based methods for automatic segmentation of prostatic zones. After 2017, almost all publications were based on deep learning with convolutional neural networks (CNN) (72%, 24/33). Common architectures such as U-net [11] have been extensively used, with modification and fine tuning of existing models, allowing either improved accuracy of classical networks or reduced memory and storage requirements.

Dice coefficient (DSC) and Hausdorff distance [47] were commonly used metrics. Almost all authors found inferior results for PZ than WG, CG or TZ segmentation, attributing this to the more complex shape and structure of PZ, especially within the anterior bundles. Eleven authors subsequently stratified their DSC results based on prostate height, with various methods:in three equal parts [13], in 25% apex, 50% mid gland and 25% base [39] in 30%, 40% and 30%, respectively [31]. Five authors did not provide any details on how they divided the volume.

These results as well as the remaining metrics are summarized in Table 5.

Table 5 Overview of segmentation methods with performance based on DSC. Number of articles reporting stratification by gland height, and reporting pre- or post-processing steps
Full size table

Discussion

Our systematic review highlights the high prevalence of deficiencies in methodology in the literature on automatic segmentation of prostate gland on MRI.

Since 2011, 33 studies proposed new or fine-tuned existing approaches for automatic prostatic zonal segmentation. Many studies are hampered by issues with limitation of the dataset used in the model, methodological mistakes, poor reproducibility, and biases in study design. Most studies focused on achieving the best accuracy for their algorithms, sometimes putting aside validity and applicability in clinical practice. Indeed, only two articles presented with an overall low risk of bias.

The common limitations concerned datasets used for the model development, definition of the ground truth for evaluation of the model and strategies used for model evaluation.

Regarding the datasets used, some are private, and some are public open source. For private databases, advanced technical characteristics of images (e.g., imaging sequence, field of view, noise) used and patient’s inclusion and exclusion criteria were poorly or not described. Most databases lacked representability of patients’ variability as prostate volume, prostate tissue heterogeneity, prostatic pathology as PCa or benign hypertropia. Open-source prostate MRI databases also have several limitations such as selection bias, limited annotations, low-resolution images, unclear terminology, lack of demographic statistics and of precise histologic data.

This can have a direct impact on the generalizability of the model developed. Indeed, it has been shown for example that prostate morphological differences contribute to segmentation variability: Montagne et al. [48], showed that the smaller the prostate volume was, the higher the variability was; several authors [18, 39, 43] found poorer performance of their model applied on special cases such as history of trans-urethral-resection of prostate (TURP), while most databases lacked representativity of patients variability.

Even though it is tedious and time-consuming, reference segmentation should require at least two trained readers because inter- and intra-rater variability can be significant. Quality of images (slice thickness, partial volume artifacts), apex or base location [48, 49] or prostate morphological differences [48] have been shown to decrease accuracy of segmentation. Meyer et al. [34] showed that training on segmentation obtained by a single reader introduced bias into the training data. Indeed, performance was higher when obtained from the expert who created the training data in comparison with evaluation against other expert segmentation. Aldoj et al. [39] emphasized the need for finely annotated sets as they improved overall performances of their algorithms, showing the greater importance of well annotated databases compared to large and coarsely annotated databases.

Quality of the resulting auto segmentation is evaluated against the corresponding reference segmentation, so called the ground truth. The main approach is manual delineation of the prostate zones performed by human experts. We found a great heterogeneity on the segmentation protocols and terminology used. Eighteen different types of prostate delineation were found; each anatomical zone was segmented directly or obtained by subtraction from one region to another (resulting in CZ, AFMS and PZ, which can be obtained either by delineation or by subtraction of WG and TZ). Terminology used was extremely variable from one study to another and did not always respect the one used and referenced in the PI-RADS [3, 50] (for example, use of “central gland” instead of CZ or TZ).

Number of readers, level of expertise, inter- and intra-variability evaluation were mostly absent, limiting the generalizability of the developed models due to inter-observer variability. Only 2/33 studies [10, 23] used blinded reading for ground truth. Nonetheless, prostate segmentation is a very challenging task. The prostate gland usually has fuzzy boundaries. Pixel intensities are heterogeneous both inside and outside the prostate, and contrasts and pixel intensities are very similar for prostate and non-prostate regions. The manual delineation of the prostate zones is therefore limited by the subjective interpretation of the organ boundaries. Becker et al. [49] found in a multi-reader study a higher variability at the extreme part of the gland (apex and base) and for the TZ delineation. Similar results were found by Padgett et al. [8] who found a difference of DSC from 0.88 to 0.81 for WG and TZ. Meyer et al. [34] showed that training on segmentation obtained by a single reader introduced bias into the training data.

Strategies used for model evaluation were limited by the lack of external validation Only 7 studies [6, 9, 14, 30, 33, 36, 37] used both private and public data to evaluate their model. The absence of an external testing dataset is a critical limitation to the clinical applicability of the developed models. Data augmentation and transfer learning were also used to help addressing this issue [6, 14,15,16, 29, 31, 33, 35,36,37,38,39,40,41, 43, 51]. It is important to note that some bias cannot be balanced-out by increasing the sample size by data augmentation or repetition of training. For example, data augmentation of a dataset constituted without prostate cancer patients cannot decrease risk of bias induced by the more homogeneous contours it provides.

Even without data augmentation, MRI images contains wide heterogeneity and most of the times pre-processing steps involving intensity normalization or noise reduction to remove confounding features and improve image quality are necessary [52]. Some authors [6, 13, 15, 31, 33, 35, 51] also reported post-processing. Not reporting some of the pre- or post-processing steps can affect reproducibility and sufficient detail enables readers to determine the quality and generalizability of the work. While several checklists can be used such as those from Enhancing the Quality and Transparency Of health Research (EQUATOR) Network guidelines [53], the use of the recently published Checklist for Artificial Intelligence in Medical Imaging [22] would be helpful to lower risk of bias of ongoing work.

In the future, there is a need for well-sampled databases including large number of representative cases for the anatomical variability of the prostate gland and technical specificities (2D T2 versus 3D T2, slice thickness, FOV, vendors) to account for the anatomical, disease related, acquisition related variabilities, with a multi-readers segmentations and a well-defined delineation guideline of the prostate (as it is already done for example in organs at risk for radiotherapy planning [54]).

Constitution of quality database should be based on latest PI-RADS recommendations, by associating quality criteria such as the consensual quality requirements ESUR/ESUI [55] or Prostate Imaging Quality (PI-QUAL) [56] score to guarantee essential image quality for zonal segmentation and tumor detection.

The main limitation of this review is the absence of details of technical information used; each study making its own contribution for networks with countless hyperparameters, sometimes without enough details to be gathered. This precluded us from comparing models’ accuracy without bias.

Some other relevant papers also could be missing because of incongruences between search terms, article keywords, or indexing in the databases, such as for conference proceedings papers. In particular, databases such as ArXiv were not searched as it also provides access to preprints, without peer review.

Conclusion

This review systematically synthesizes published automatic prostate zonal segmentation methods using MRI. We found that no papers in the literature currently have both sufficiently documented datasets selection and segmentation criteria and enough external validation.

This underlines the critical need for higher quality datasets, a documented reproducible method and terminology for zonal segmentation and sufficient external dataset to develop the best quality methods free from biases: an essential step for future development of automatic detection of prostate cancer.