Advanced AI-Based Multi-Stage Deep Learning Architecture for Robust Image Segmentation and Comprehensive Classification of Interstitial Lung Disorders in Next-Generation Medical Diagnostics

Abstract

Current trends in interstitial lung disease research focus on early detection, precise diagnosis, and thorough characterization to optimize patient outcomes. Challenges persist because of heterogeneous patterns of disease, the scarcity of annotated datasets, and variability in image quality, resulting in inconsistent diagnosis and prognosis. To overcome this problem, in this manuscript proposes image segmentation and comprehensive classification of interstitial lung disorders using Xception Spiking Fractional Neural Network (ISCC-ILD-XSFNN) is proposed. The proposed model integrates Fuzzy Consensus Cubature Information Filtering (FCCIF), Structured Doubly Stochastic Graph-based Clustering (SDSGC), Self-Modulating Convolutional Neural Networks (SMCNet), and Xception Spiking Fractional Neural Networks (XSFNN) into a cohesive architecture. Initially, lung images from the MedGIFT database are preprocessed using FCCIF for normalization and noise reduction. The normalized images are segmented through SDSGC to identify affected lung regions. Subsequently, SMCNet extracts discriminative multi-scale features from segmented images. Finally, XSFNN classifies interstitial lung diseases into consolidation, ground glass, fibrosis, micronodules, emphysema, and normal classes. Experimental results demonstrate better and balanced classification performance, achieving 99.12% accuracy, 99.23% precision, 99.07% recall and 99.14% F1-score. Also, the per-class metrics remain consistently high across all six classes, outperforming existing state-of-the-art approaches. These results indicate that the proposed framework maintains balanced predictive performance across different ILD categories.