MULTI-TASK DEEP LEARNING FOR KNEE MRI PATHOLOGY CLASSIFICATION
Abstract
Magnetic resonance imaging (MRI) is the primary non-invasive modality for diagnosing internal knee derangement, but manual reading of multi-planar examinations is time-consuming and subject to inter-observer variability, particularly for anterior cruciate ligament (ACL) and meniscus tears. Most deep learning approaches train a separate single-task model for each pathology, which limits the use of shared anatomical representations and complicates deployment. This study develops and evaluates a unified multi-task, multi-view framework that classifies three knee pathologies from multi-plane MRI in a single forward pass. Using the public part of the MRNet dataset (1,250 examinations with axial, coronal and sagittal series), we designed Multi-View Multi-Task MRNet (MVMT-MRNet), in which one ResNet50 backbone shared by all planes encodes each plane, slice features are max-pooled, and a late-fusion module feeds three classification heads. The 120 public validation examinations served as the held-out test set. Training used augmentation, class-weighted binary cross-entropy, mixed precision and gradient accumulation over eight examinations with a learning rate of 8×10⁻⁵, selected on validation data. On the test set the multi-task model reached AUCs of 0.844 (95% CI 0.752–0.922) for abnormality, 0.914 (0.857–0.959) for ACL tear and 0.664 (0.567–0.757) for meniscus tear, a mean of 0.807 (0.749–0.860). Three single-task models of identical architecture reached a mean AUC of 0.853 (0.802–0.897); the paired difference of −0.046 (−0.115 to +0.015) includes zero, but the multi-task model was significantly worse for meniscus tear (−0.126, −0.231 to −0.020). Averaging gradients over eight examinations helped only when the learning rate was scaled accordingly. A single shared model (25.3 M parameters instead of 75.8 M, one forward pass instead of three), trainable on one consumer GPU, thus matches three single-task models on mean AUC, but parameter sharing redistributes accuracy across tasks rather than adding it. External validation is required before clinical use. Code: https://github.com/ Amanekkkk/MRNet_multihead.