Liver Lesion Classification from Multi-Labeled Volumetric CT Scans Using HepatoSeqNet

Abstract

Early detection and treatment are essential for liver cancer because early-stage disease often lacks prominent symptoms, which can lead to late diagnosis. This paper introduces HepatoSeqNet, a hybrid deep learning framework that combines a ResNet-18 backbone for extracting spatial features from individual slices with a Long ShortTerm Memory (LSTM) network for modelling long-term dependencies between slices. This framework performs multi-label classification across five classes: Normal Liver Tissue, Liver Cyst, Cavernous Hemangioma, Hepatic Metastases, and Hepatocellular Carcinoma. The final configuration uniformly samples up to 60 axial slices from each volume, applies data augmentation, and trains using Binary Cross-Entropy loss with mixed-precision gradient accumulation. Five-fold evaluation yields an exact-match accuracy of 73.33%±5.16%, macro precision of 76.95%± 6.17%, macro recall of 72.06%±2.70%, and macro F1-score of 73.98%± 2.71%.