Novel attention-enhanced Multi-Task Deep learning for knee osteoarthritis (KOA) grading and localization in X-ray imaging of basketball players

Abstract

Objective

This study aims to develop and evaluate a novel attention-enhanced multi-task deep learning framework designed to automatically detect, grade, and localize the severity of knee osteoarthritis (KOA) in basketball players through X-ray imaging. The framework utilizes the Kellgren-Lawrence (KL) grading system, categorizing KOA severity into grades 1 to 4, to enhance diagnostic accuracy, reliability, and clinical applicability.

Materials and methods

The dataset consisted of 2135 knee X-ray images from basketball players, labeled with KL grades ranging from 1 (Normal) to 4 (Severe). The images were preprocessed using normalization and data augmentation techniques, such as rotations, flips, and intensity adjustments, to increase diversity and enhance model performance. YOLOv11, built on the YOLO framework, incorporated architectural improvements like enhanced feature pyramids, adaptive anchor thresholds, and attention mechanisms. These features allowed the model to simultaneously detect, classify, and locate pathological features. The model was trained for 500 epochs using a composite loss function that balanced the objectives of detection, classification, and localization. Hyperparameter tuning was used to optimize the learning rate, batch size, and anchor thresholds. The model's performance was evaluated using metrics like mean Average Precision (mAP), Intersection over Union (IoU), and classification measures such as precision, recall, accuracy, and F1 score on training, validation, and testing datasets.

Results

YOLOv11 achieved a final mAP of 97.23 % and IoU of 96.54 %, outperforming YOLOv10 (mAP: 95.58 %, IoU: 92.16 %) and YOLOv9 (mAP: 93.14 %, IoU: 90.12 %). The model also showed excellent classification performance, achieving an accuracy of 97.70 % on the training dataset, 96.22 % on the validation dataset, and 95.59 % on the testing dataset. It effectively identified KOA-related features with minimal errors, as shown by confusion matrices and t-SNE visualizations, which displayed clear clustering of KL grades.

Conclusions

YOLOv11 marks a significant step forward in the automated assessment of KOA, offering high precision, accuracy, and reliable localization. Its success with basketball players indicates its potential for use in other athletic populations and broader clinical applications.