An explainable hybrid SMPR-Net-XGBoost framework for automated lupus nephritis detection using medical image analysis.
Authors
Affiliations (2)
Affiliations (2)
- Department of Computer Science and Engineering, Siddhartha Academy of Higher Education, Deemed to be University, Vijayawada, 520007, Andhra Pradesh, India. [email protected].
- Department of Computer Science and Engineering, Siddhartha Academy of Higher Education, Deemed to be University, Vijayawada, 520007, Andhra Pradesh, India.
Abstract
Lupus Nephritis (LN) is among the worst renal manifestations of Systemic Lupus Erythematosus (SLE) that may lead to irreversible damage to the kidneys without early diagnosis. The traditional diagnostic methods, especially renal biopsy, are invasive, expensive, and less suitable for frequent monitoring, which places a high demand on the reliability of computer-assisted diagnostic methods. This paper suggests a hybrid Deep Learning (DL) model that can explain the process of automated LN detection based on medical image analysis. The proposed framework uses Data-Adaptive Gaussian Average Filtering (DAGAF) to reduce the noise and retain diagnostically important renal structural and textural data. The dual strategy is used to extract features based on Adaptive Residual Exemplars Local Binary Pattern (ARE-LBP) to provide robust handcrafted texture description and DL of features using a novel Self-Modulated Convolutional Neural Network with Partial Multi-Scale Channel Attention and Residual Dual Pooling (SMPR-Net) model. In order to increase the robustness of classification and minimize misclassification in complicated decisions, high-level deep features obtained because of the proposed network are classified again with Extreme Gradient Boosting (XGBoost). The experimental assessment of the Lupus Class dataset illustrates that the proposed SMPR-Net-XGBoost model has a high performance, with the highest classification accuracy of 97.18, an F1 score of 95.87, and ROC-AUC of 0.99 with an 80 percent training split, which is significantly better than the traditional Machine Learning (ML) and the traditional DL models. Moreover, SHAP and LIME offer clear and clinically understandable explanations of the model predictions, increasing the credibility and applicability of the model in clinical environments.