Explainable AD classification using integrated quantum-inspired deep neural and transformer models
TL;DR
An explainable and robust classification framework that will be able to retain its performance under in-distribution and blind-testing settings and validating its generalization ability of domain shift is proposed.
Abstract
The classification of Alzheimer’s disease (AD) is often not consistent and reliable due to heterogeneous neurodegenerative patterns and the non-generalization of various imaging sources. This study proposes an explainable and robust classification framework that will be able to retain its performance under in-distribution and blind-testing settings. The study uses 4 publicly available brain MRI datasets covering several stages of AD including non-demented, mild cognitive impairment and progressive Alzheimer’s stages. The datasets consist of balanced and augmented scans, and clinically heterogeneous scans were acquired using different imaging and acquisition techniques. A dedicated independent dataset is used for blind evaluation to check robustness against unseen-source domain shifts. An explainable hybrid framework, named Tri-Fusion-ADNet, is proposed in this work. It consists of EfficientNet-V2-S for localized neuro-anatomical feature extraction, Swin TransformerV2 for modelling long-range contextual dependencies and a quantum-inspired variational neural network (QI-VNN) for enhanced interaction of higher-order features. Before the classification, fusion has been done on feature levels. The reliability of the model is checked through 10-fold cross-validation on various datasets. Further, qualitative interpretability has been supported through visualization using Grad-CAM.The proposed Tri-Fusion-ADNet performed well in all data sets. The testing accuracy of the model was 98.42% (MCC = 0.979) on the balanced augmented dataset (D-I). The accuracy was 97.86% (MCC = 0.973) on augmented dataset (D-II). The clinically heterogeneous dataset (D-III) gave 96.21% (MCC = 0.949). Also, 91.18 when trained on D-I and tested on D-II, and 91.64 when trained on D-II and tested on D-I, cross-dataset testing without adaptation yields very high accuracies. In addition, during blind cross-dataset evaluation on an independent unseen dataset (D-IV), the framework maintained a robust accuracy of 91.22%, validating its generalization ability of domain shift.