Artificial Intelligence for Molecular Subtyping in Unresectable Gallbladder Cancer: A Proof-of-Concept Study for CT-based HER2 Status Prediction.
Authors
Affiliations (12)
Affiliations (12)
- Department of Radiodiagnosis, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Computer Sciences, Indian Institute of Technology, New Delhi, India, 110016.
- Department of Clinical Hematology and Medical Oncology, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of GI Surgery, HBP and Liver Transplantation, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of General Surgery, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Cytology and Gynaecological Pathology, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Histopathology, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Radiotherapy, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Nuclear Medicine, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Department of Medical Gastroenterology, Postgraduate Institute of Medical Education and Research, Chandigarh, 160012, India.
- Institute of Nuclear Medicine, University College London, UK.
- University of Alabama, Birmingham, AL, USA.
Abstract
Human epidermal growth factor receptor 2 (HER2) overexpression is a critical therapeutic target in gallbladder cancer (GBC), but detection requires invasive sampling. We developed a fully automated computed tomography (CT)-based framework to noninvasively predict HER2 status in unresectable GBC. This single-center retrospective-prospective study included 213 pathologically proven unresectable GBC patients (retrospective training cohort, n = 143; prospective test cohort, n = 70) with 42 (19.7%) HER2-positive cases. Our two-stage framework comprised: (1) automated tumor detection and segmentation using Grounding DINO-MedSAM, and (2) HER2 classification through three parallel approaches: clinical models, radiomics models, and deep learning (DL) models with attention mechanisms. The detection and segmentation performance was assessed using Dice and intersection over union (IoU). HER2 classification performance was evaluated using sensitivity, specificity, area under the receiver operating characteristic curve (AUROC), and F1 scores. The segmentation pipeline achieved a mean Dice score of 0.62 and IoU of 0.53 on the test set. Clinical models showed limited performance (random forest: sensitivity 0.222, specificity 0.788, AUROC 0.576 on the test set). Radiomics models performed better than clinical models (random forest: sensitivity 0.824, specificity 0.471, AUROC 0.659 on the test set). DL models demonstrated improvement over clinical and radiomics models. In the test set, the Swin Transformer DL model achieved a sensitivity of 0.729 (95% confidence interval [CI], 0.670-0.800), specificity of 0.857 (95% CI, 0.780-0.930), AUROC of 0.792 (95% CI, 0.733-0.861), and F1-score of 0.769 (95% CI, 0.711-0.835). The DenseNet DL model achieved a sensitivity of 0.750 (95% CI, 0.690-0.810), specificity of 0.791 (95% CI, 0.627-0.719), AUROC of 0.763 (95% CI, 0.697-0.799), and F1-score of 0.762 (95% CI, 0.679-0.801) on the test set. There was no significant difference in the classification performance of radiomic or DL models using ground truth versus automated segmentations. This study establishes a fully automated CT-based deep learning pipeline for HER2 status prediction in unresectable GBC. Validation in larger, multi-institutional cohorts is warranted.