Deep Learning-Driven Diagnosis CNN-LSTM Models for Accurate Grading of Knee Osteoarthritis

Musculoskeletal disorder knee osteoarthritis (OA) is very prevalent and impacts quality of life demanding accurate grading to inform treatment planning. This study presents a novel deep learning approach based on combination of Convolutional Neural Networks (CNN) and Long Short Term Memory (LSTM) networks to enhance the accuracy of OA grading. In this, we integrate spatial feature extraction by a ResNet50 backbone and temporal sequence modeling by LSTM which is able to capture both the structural and temporal patterns from radiographic images. Experimental evaluations using the OAI and KAG datasets together with radiographs images classified according to the Kellgren-Lawrence (KL) scale are presented. Using the proposed CNN-LSTM model, we achieve accuracy of 92.3%, outperforming previous traditional models (ResNet50 achieves 87.5%; DenseNet-LSTM achieves 89.8%). On the other hand, the model also did better in terms of precision (91.5%), recall (91.8%) and F1-score (91.6%) which shows that it worked well in terms of balance between false positives and false negatives. The robustness of the model was further confirmed via receiver operating characteristic (ROC) analysis with an area under the curve (AUC) greater than 0.90 across all KL grades. Through varying interpretability, the Grad-CAM visualizations helped validate the results by identifying key join areas of relevance to the OA task. The results show that the proposed CNN-LSTM framework is an accurate, reliable and interpretable solution for automated OA grading. Future work will include improving generalization of the model by augmenting the dataset with multi-modal data from many different patient populations.