An Explainable Deep Hybrid Architecture DMI Net for Knee and Hip Implant Status Classification in X-ray Imaging

Abstract

Objective: Complications in knee and hip joint replacements are implant-related current clinical concerns, and the delay or inaccuracy in diagnosis of these complications may result in serious patient outcomes and revision surgeries. To overcome these issues, this paper proposes a new type of hybrid Deep Multiscale Integration Network (DMI-Net) that can be used to accurately assess the condition of the implants using knee and hip X-ray datasets.
Method: The proposed structure includes a full preprocessing pipeline, image resizing, contrast enhancement using contrast-limited adaptive histogram equalization (CLAHE), normalization, and Gaussian filtering, to enhance visual quality and highlight important structural details. A lightweight parallel depthwise separable convolutional neural network (PD-CNN) is used to effectively extract multiscale features with minimal computational complexity. As an extension of discriminative capability, the architecture combines several high-performance transfer learning models through feature stacking and concatenation to form a strong hybrid system. To gain further insight and confirm feature separability, dimensionality reduction algorithms like PCA, LDA, UMAP, and t-SNE are employed. Also, SHAP and Grad-CAM approaches are used to enhance the interpretability of models, which allows visualizing of decision-relevant regions of the model and enhances clinical trust.
Results: The proposed DMI-Net has shown excellent performance when compared to some of the established models, with an accuracy of 98.92, precision of 98.70, recall of 98.60, F1 score of 98.65 and AUC of 98.80. These results point to the efficacy and validity of the suggested approach, indicating its high potential as a supportive tool to make accurate diagnosis of the implants and well-informed clinical decisions.
Conclusion: The present study is a contribution of a novel hybrid DMI-Net framework for detecting and classifying orthopaedic conditions using combined knee and hip implant X-ray images. Overall, the suggested hybrid DMI-Net is significantly more dependable, powerful, and understandable in computer-aided diagnosis of orthopaedic imaging, particularly the evaluation of knee and hip implants.