A Multi-Task Attention-Driven Framework for Lung Infection Segmentation and Cancer Classification From HRCT Images

Lung infections and cancer are major causes of mortality worldwide, requiring accurate detection and diagnosis from high-resolution computed tomography (HRCT) images. This project develops two independent attention-driven deep learning frameworks: CPRL‑Net and TH‑Net. CPRL‑Net leverages boundary uncertainty maps from infection segmentation to enhance early-stage cancer classification, emulating radiologist workflows by focusing on lesion margins and edge characteristics. TH‑Net introduces bidirectional attention and task harmonization to independently optimize segmentation and classification tasks without relying on early transfer, using learnable scaling vectors and dynamic task weighting to balance losses automatically. Two datasets are employed: a lung cancer classification dataset and a COVID-19 lesion segmentation dataset. Preprocessing steps include resizing, normalization, and data augmentation to standardize input images and improve model performance. CPRL‑Net uses interleaved training batches and applies boundary priors during initial epochs to guide classification, while TH‑Net continuously harmonizes features across tasks using mutual attention and dynamic loss adjustment. Both models provide interpretable outputs, including segmentation masks, attention maps, and classification probabilities, enabling visual understanding of lesion location and severity. Experimental results indicate that CPRL‑Net excels in capturing boundary-sensitive features, improving classification accuracy, while TH‑Net efficiently balances segmentation and classification tasks for robust performance. This project demonstrates how task-specific attention mechanisms and boundary-aware feature learning can improve lung disease detection and support clinical decision-making.

Keywords: HRCT, lung cancer, COVID‑19, segmentation, classification, deep learning, attention mechanism, CPRL‑Net, TH‑Net, boundary uncertainty.