Fully Automated Measurement of the Insall-Salvati Ratio with Artificial Intelligence

Abstract

Patella alta (PA) and patella baja (PB) affect 1–2% of the world population, but are often underreported, leading to potential complications like osteoarthritis. The Insall-Salvati ratio (ISR) is commonly used to diagnose patellar height abnormalities. Artificial intelligence (AI) keypoint models show promising accuracy in measuring and detecting these abnormalities.

An AI keypoint model is developed and validated to study the Insall-Salvati ratio on a random population sample of lateral knee radiographs. A keypoint model was trained and internally validated with 689 lateral knee radiographs from five sites in a multi-hospital urban healthcare system after IRB approval. A total of 116 lateral knee radiographs from a sixth site were used for external validation. Distance error (mm), Pearson correlation, and Bland–Altman plots were used to evaluate model performance. On a random sample of 2647 different lateral knee radiographs, mean and standard deviation were used to calculate the normal distribution of ISR. A keypoint detection model had mean distance error of 2.57 ± 2.44 mm on internal validation data and 2.73 ± 2.86 mm on external validation data. Pearson correlation between labeled and predicted Insall-Salvati ratios was 0.82 [95% CI 0.76–0.86] on internal validation and 0.75 [0.66–0.82] on external validation. For the population sample of 2647 patients, there was mean ISR of 1.11 ± 0.21. Patellar height abnormalities were underreported in radiology reports from the population sample. AI keypoint models consistently measure ISR on knee radiographs. Future models can enable radiologists to study musculoskeletal measurements on larger population samples and enhance our understanding of normal and abnormal ranges.

Appendix

Neural Network Architecture

The Mask R-CNN architecture was used, with standard modifications for keypoint detection [35]. This network is based on a Convolutional Neural Network (CNN) with additional components for detecting, localizing, and classifying objects on an image. The complete model architecture is detailed in Figure S1. Briefly, an input image is first processed by a base CNN, which consists of convolutional and max-pooling layers, to produce feature maps. After the last convolutional layer, a regional proposal network is trained to propose candidate keypoint detections on the image on the basis of a fixed set of anchors on each position of the feature maps. The location and size of each anchor are fine-tuned with bounding-box regression. Candidate keypoints are filtered by performing nonmaximum suppression which removes redundant overlapping candidate regions. Region proposals are then sent to a region-of-interest pooling layer, which resamples the feature maps inside each proposal and is fed to another branch of the network that predicts the confidence scores. The base CNN used in our model was ConvNeXt_Large, which was set with pretrained weights from ImageNet, a large-scale object detection, segmentation, and captioning dataset [36, 37]. Ultimately, the keypoint model outputs three sets of x and y coordinates (keypoints) for each image: one set for the superior aspect of the patella, one set for the inferior aspect of the patella, and a third for the tibial tubercle, where the patellar tendon attaches. This network was implemented via the PyTorch v2.0.0 software package with Python 3.6.5, and training was done on Nvidia K80 and Nvidia P100 GPUs (Nvidia, Inc, Santa Clara, CA).