Summary

A trained machine learning algorithm demonstrated capabilities to both localize the ACL on MRI and predict extent of injury..

Abstract

Objectives:
Anterior cruciate ligament (ACL) injuries are among the most common surgical indications in orthopaedic sports medicine. Assessing the extent of damage in ACL injury is a challenging task usually performed by interpreting magnetic resonance imaging (MRI). Machine learning (ML) modalities can potentially act as a diagnostic adjuncts in this process. Our team explored the feasibility of training a deep learning model to perform this task. The purpose of this study was to determine if a neural network trained on an annotated dataset of knee MRIs can (1) accurately localize the ACL from lateral views of the knee and (2) accurately predict the extent of the ACL injury, if any.

Methods

3,059 MRI studies containing lateral views of the knee were annotated with a tightly cropped bounding box around the ACL and labeled as (1) Healthy, (2) Partially-injured, or (3) Completely-ruptured. The annotated dataset was then uniformly resized to 256x256 pixels and randomly split in into training, validation, and test data subsets in a 70:20:10 ratio, respectively. The training dataset was augmented using randomly-applied affine image transformations to generate 3 training samples per original image. These transformations included horizontal flipping, rotating between -2 and +2 degrees, and varying image exposure between -3% and +3%. The final, augmented dataset included 6.3k training images, 612 validation images, and 306 test images. This dataset was used to train an object detection neural network for 150 epochs, tracking its progression against validation data at the end of each epoch. Loss functions were used to track the model’s error rates in generating bounding boxes and qualitative predictions of ACL integrity. Mean average precision (mAP), overall precision, and recall rates were recorded throughout the training process.

Results

The model’s average precision by on validation data was 97% on healthy ACLs, 73% on partial ACL injuries, and 89% on complete ACL ruptures. The model’s average precision by on test data was 98% on healthy ACLs, 72% on partial injuries, and 84% on complete ruptures. Overall average precision was 86% on validation data and 85% on test data. The trained model had a mAP of 86.2%, precision of 78.4%, and recall of 84.7%.

Conclusions

Despite the difficulty of the tasks, the model performed reasonably well in both demarcating ACL location on imaging and qualitatively assessing its structural integrity. This study supports the promising future of machine learning applications in healthcare and prompts further exploration of modalities that leverage these capabilities. The results of this study may serve as a benchmark for iterative model improvement and comparing alternative machine learning approaches.