Automated Recognition of healthy Anterior Cruciate Ligament in Sagittal MR images using Lightweight Deep Learning

Anterior cruciate ligament (ACL) tears are very common among athletes. The success of enhanced ACL injury therapy hinges on accurate and cost-effective detection. Deep learning-based techniques have recently dominated ACL injury detection in MRI research. The goal of this study is to develop a robust and lightweight deep learning pipeline for identifying ACL in 3D MRI data of healthy knees. Specifically, we aim at finding the slices in the sagittal plane where the ACL is present. This could be utilized by clinicians for further evaluation. To this end, we build and test an advanced pipeline that relies on the newest object detection network, YOLOv5-Nano. We go on to compare our model to other pipelines that rely on YOLOv5-xlarge, YOLOX-small and YOLOX-nano. YOLOv5-nano is shown to be the best performer, obtaining the highest overall mAP@0.5 performance (0.9727) on augmented data, while at the same time having the smallest model size (3.7 MB). Conclusive object detection is a key step in identifying damage. YOLOv5-nano offers a great solution towards achieving robust object detection healthcare systems that will permit local processing by devices with limited computational resources.