Hierarchical diffusion-guided consistency learning framework for knee anterior cruciate ligament classification

Anterior cruciate ligament (ACL) classification in knee magnetic resonance imaging (MRI) is challenging due to diagnostic evidence being dispersed across multiple slices and anatomical supervision often being sparse. We propose a three-stage hierarchical diffusion-guided consistency learning framework that integrates global multi-class estimation, anchor-versus-residual separation, and within-residual refinement. Each private case uses one to four expert-annotated ACL masks to supervise structural latent representations; finite-step diffusion perturbation-recovery establishes consistent and reliable training targets, while ROI Collaborative Interaction (RCI) injects localized evidence into staged representations, and reliability-aware soft integration combines global and hierarchical predictions. No iterative reverse diffusion sampling is required in the inference phase. The framework was evaluated on 300 private cases and four public datasets using dataset-specific five-fold cross-validation and six external-transfer settings. On the private cohort, the method achieved an Accuracy of 89.79 ± 0.45% and a Macro-F1 of 89.13 ± 0.27%; on the KneeMRI dataset, these were 87.63 ± 1.08% and 83.11 ± 1.16%, respectively, representing improvements of 6.01 and 2.25 percentage points over the strongest baselines, which remained significant after Holm correction. Across the five datasets, the method achieved the highest Accuracy on MRNet, KneeMRI, and ACL-PCL, and the highest Macro-F1 on the private cohort and KneeMRI, although it did not dominate in all metrics. External-transfer Accuracy and Macro-F1 ranged from 73.82% to 85.25% and from 71.28% to 84.46%, respectively. These results indicate the dataset-dependent efficacy of the proposed method under sparse structural supervision and highlight the necessity of multicenter validation and domain adaptation research.