COMPARISON OF CNN ARCHITECTURES FOR AUTOMATED ACL INJURY CLASSIFICATION FROM KNEE MRI

B. Bekenuly, A. Baimakhanova, Vinera Baiturganova, Dinara Bolatbekova

2026

Closed access Metadata (CSL-JSON)
DOI 10.32014/2026.2518-1726.455
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Relevance. Anterior cruciate ligament (ACL) tears are among the most clinically significant knee injuries, and delayed or inaccurate diagnosis is associated with progressive joint degeneration. Magnetic resonance imaging (MRI) is the primary non-invasive modality for assessing ligament integrity, but its interpretation is time-consuming, requires considerable expertise and remains vulnerable to fatigue and cognitive bias, which motivates the development of automated decision-support tools. Aim. The study systematically evaluates and compares four convolutional neural network (CNN) architectures trained from scratch for three-class ACL injury classification (healthy, partial tear, complete tear). Methods. A publicly available dataset of 736 sagittal-plane knee MRI volumes (a subset of the KneeMRI dataset) was used. Each 3D volume was reduced to a 2D image by extracting the middle sagittal slice and resizing it to 128 × 128 pixels. ResNet-18, MobileNetV2, EfficientNet-B0 and MobileNetV4_small were trained end-to-end under a unified configuration with the AdamW optimizer, cosine annealing, cross-entropy loss and weighted random sampling to compensate for class imbalance. Each model was trained with five independent random seeds, and performance is reported as mean ± standard deviation of weighted precision, recall, F1-score and AUC. Results. ResNet-18 achieved the best performance (F1-score 0.6637 ± 0.0351; AUC 0.6622 ± 0.0136), followed by EfficientNet-B0, MobileNetV4_small, and MobileNetV2. Practical significance. The study provides a reproducible benchmark and practical guidance for architecture selection in small-scale medical imaging tasks, showing that, when training from scratch, the input representation and the availability of pre-trained features matter more than the choice of backbone, while clinical reliability will require 3D volumetric context, pre-training and external validation.