Publication

Abstract : Knee osteoarthritis (KOA) is a progressive joint disorder that results in compromised mobility and quality of life worldwide. Early detection and accurate assessment are critical for halting the process of severe degeneration and improving treatment plans. The first three challenges to be tackled in deep learning approaches included finding low-accuracy detections for early grades; they were noise-prone and exhibited limited generalization in other imaging modalities. This situation led us to propose a unified system involving segmentation and classification in a single pipeline. The attention-based, multi-resolution segmentation architecture that we proposed, namely MtRA-UNet with shape-reconstruction loss, was mobilized to lend an anatomy-like structure-consistency view to bone and cartilage structures in the evaluation of MRIs. Hybrid Classifier that fused EffResNet-ViT for the retrieval of local features by means of texture information and global features by contextual relationships for the Kellgren-Lawrence (KL) grading was utilized to grade the radiographs. The classification also involved feature selection performed via the Artificial Hummingbird Algorithm (AHA), while hyperparameter tuning optimization was adaptively performed by the Meerkat Clan Algorithm (MCA), seeking the best configuration of the model without a trade-off on generalization. The proposed framework was validated on the OAI-ZIB MRI dataset of 507 scans and KL-graded X-ray dataset of 1650 images. Experimental results demonstrate its superiority, with a classification accuracy of 99.15% and a mean Dice similarity coefficient of 98.20% for segmentation tasks. A comparison with some of the state-of-the-art recent studies showed consistent superiority over existing methods based on CNN, transformer, and ensembles.