Diagnosis of SLAP lesions on shoulder MRI using a 2.5D deep learning and ensemble learning framework

TL;DR

This paper presents an AI model that helps diagnose SLAP shoulder injuries using MRI scans, combining deep learning and ensemble methods to improve accuracy.

Contribution

A novel 2.5D deep learning and ensemble learning framework for automated SLAP lesion diagnosis on shoulder MRI is introduced.

Findings

The XGBoost model achieved the highest AUC (0.754) and sensitivity (0.933) for SLAP lesion detection.

The central MRI slice was identified as the most contributory feature for diagnosis.

The model's focus on the biceps-labral complex aligns with known SLAP pathology anatomy.

Abstract

Superior labrum anterior and posterior (SLAP) lesions are a common cause of shoulder pain and instability. Developing accurate, non-invasive diagnostic tools is essential to support clinical decision-making for SLAP lesions. This study aimed to establish an automated diagnostic model for SLAP lesions using a 2.5D deep learning framework combined with ensemble learning and to evaluate its clinical utility.

In this retrospective study, 185 patients who underwent shoulder arthroscopy between January 2019 and September 2025 were included (91 SLAP lesions, 94 controls). Preoperative shoulder magnetic resonance imaging (MRI) data were analysed. Images from three consecutive slices, centred on the maximal region of interest (ROI), were processed using a Wide_ResNet101_2 network pre-trained on ImageNet for deep feature extraction and probability prediction. A decision-level fusion strategy integrated the predicted probabilities from all three layers as input features for three ensemble classifiers: AdaBoost, Random Forest, and XGBoost. Model performance was assessed with accuracy, area under the receiver operating characteristic curve (AUC), sensitivity, specificity, precision, and F1-score. The DeLong test and integrated discrimination improvement (IDI) were used to compare models.

All ensemble models exhibited robust diagnostic performance. On the test set, the XGBoost model achieved the highest AUC (0.754) and sensitivity (0.933), though specificity was moderate (0.538). The Random Forest model yielded an AUC of 0.745, while the AdaBoost model achieved an AUC of 0.731. F1-scores ranged from 0.75 to 0.80. There were no statistically significant differences in AUC among the models. Feature importance analysis highlighted the central MRI slice as most contributory. Model interpretability assessments showed that the network focused predominantly on the biceps-labral complex, which is anatomically consistent with SLAP pathology.

The proposed automated diagnostic model, utilising a 2.5D deep learning and ensemble approach, demonstrated favourable diagnostic performance and clinical applicability for SLAP lesion detection on shoulder MRI. Among the ensemble strategies, the XGBoost model provided the highest sensitivity, rendering it particularly suitable as a clinical decision-support tool. The multi-slice information fusion framework substantially improved diagnostic accuracy, supporting its potential as a novel artificial intelligence solution to assist radiologists in diagnosing shoulder labral injuries.

Figures