Boundary-aware multi-scale residual learning with hybrid radiomic–deep fusion for lung tumor segmentation and classification

Abstract

Accurate lung tumor segmentation and classification from computed tomography (CT) images remain challenging because of heterogeneous tumor appearance, indistinct boundaries, and inter-class similarity. This study presents BAMR-HF, an integrated framework combining boundary-aware segmentation, hybrid radiomic–deep feature fusion, and context-aware classification. BAMR-Net performs tumor segmentation using multi-scale residual learning and boundary refinement, while TRFF-Net combines radiomic and deep features through learnable weighted fusion, while MTCA-Net incorporates tumor and peritumoral contextual information. The framework was evaluated on the processed Lung Cancer Segmentation Dataset with Lung-RADS Class comprising 972 CT images, with 708 images in the training partition and 264 images in the repository-defined held-out test partition. The source labels were consolidated into Mild (LR2), Moderate (LR3), and Severe (LR4A–LR4B) categories. BAMR-Net achieved a test Dice score of 0.8563. TRFF-Net and MTCA-Net achieved test accuracies of 0.9038 and 0.9327, respectively, while the complete BAMR-HF framework attained an accuracy of 0.9478, F1-score of 0.9476, AUC of 0.9872, and MCC of 0.9215. These findings support the complementary use of boundary-aware segmentation, hybrid feature representation, and contextual modeling, while broader validation on independent multi-institutional data remains necessary.