MSF-PSD: multi-scale features based panoptic segmentation and detection framework for lung disease classification

Complex anatomical variations and heterogeneous pathological patterns make it difficult to accurately segment lung images from Computed Tomography (CT) scans, especially in lungs affected by COVID-19. Conventional segmentation techniques often struggle to capture Panoptic Segmentation information and fine-grained structures across diverse clinical datasets. This study suggests a novel Multi-Scale Feature-based Panoptic Segmentation and Detection (MSF-PSD) framework for simultaneous lesion boundary detection and lung disease classification. To improve multi-scale feature representation, the suggested framework incorporates an Efficient Channel Attention (ECA) and uses EfficientNet-B7 as its foundation. For hierarchical feature fusion, a Region Proposal Network (RPN) and Region of Interest (RoI) refinement with Non-Maximum Suppression (NMS) are used after a Feature Pyramid Network (FPN). An end-to-end panoptic detection pipeline is made possible by three parallel prediction heads that carry out category classification, bounding box regression, and mask segmentation. The COVIDx-CT and SARS-CoV-2 datasets are preprocessed using Gaussian filtering and CLAHE to increase structural visibility. Supercomputing resources are utilised to guarantee scalable training and effective inference because multi-head learning and parallel feature extraction have high computational demands. Gradient-weighted Class Activation Mapping (Grad-CAM) localises discriminative pathological regions and improves model interpretability. MSF-PSD achieves 98.4% accuracy, 97.12% precision, 94.2% recall and 93.33% F1-score, demonstrating its efficacy and clinical potential for dependable automated COVID-19 screening and lung disease assessment.