An Indian Radiologist Built an AI That Reads Gallbladder Cancer Genetics from CT Scans

Gallbladder cancer is a disease that disproportionately affects India. Diagnosed late, limited treatment options, poor survival rates, and concentrated in regions where advanced molecular testing is not always accessible. It is also a disease where knowing the HER2 status of the tumor can fundamentally change the treatment approach, since anti-HER2 therapy has shown significantly improved response rates over conventional chemotherapy.

The problem: determining HER2 status typically requires an invasive tissue biopsy. In unresectable cases, where surgery is not an option, getting adequate tissue can be difficult, risky, and sometimes impossible.

Dr. Pankaj Gupta, associate professor in the Department of Radiodiagnosis at the Postgraduate Institute of Medical Education and Research (PGIMER) in Chandigarh, asked a straightforward question: can AI figure out HER2 status from a CT scan that is already being done?

The Study: 214 Patients, Multiple AI Approaches

Funded by an RSNA Research Seed Grant, the team conducted a combined retrospective and prospective study involving 214 patients with advanced gallbladder cancer. HER2 status was confirmed using standard pathology techniques, then the researchers trained and validated multiple AI approaches using portal venous-phase CT images.

They tested machine learning, radiomics-based models, and deep learning classifiers. The deep learning approach won. An automated tumor detection and segmentation framework paired with deep learning classifiers achieved sensitivity of up to 75% and specificity of up to 86% for predicting HER2 status.

Critically, the pipeline was fully automated. Minimal human input required. The AI localizes the tumor, segments it, and predicts molecular status from the same CT scan that clinicians are already reviewing for staging.

Why This Is Not Just Another AI Paper

This research is a textbook example of AI solving a real clinical problem in a context where it matters most. In resource-limited settings where tissue biopsies are insufficient, risky, or simply unavailable, a CT-based molecular prediction tool is not a convenience. It is a lifeline.

Dr. Gupta put it directly: the model can rapidly identify patients with a high probability of being HER2-positive, expediting their access to targeted therapies that could meaningfully extend survival.

From Proof-of-Concept to Government Funding

The research has already generated momentum beyond the original study. The preliminary data provided the proof-of-concept needed to secure a substantial grant from the Department of Biotechnology, Government of India, to expand the work into chemotherapy response prediction. International collaborations in hepatobiliary cancer research have followed.

This is the trajectory that Indian radiology research needs more of: identify a locally relevant clinical problem, build an AI solution using routinely available imaging, validate it rigorously, and scale it with institutional and government support.

What This Means for Radiologists

Radiogenomics, the extraction of genomic or molecular information from imaging data, is moving from academic curiosity to clinical tool. This study demonstrates that routine CT scans contain molecular signatures that deep learning can decode.

For radiologists in India and across South Asia, this is especially relevant. Gallbladder cancer is endemic in the Gangetic belt. The imaging infrastructure exists. What was missing was the analytical layer that connects imaging patterns to molecular biology. That gap is closing.

The radiology report of the future may not just describe what the tumor looks like. It may tell the oncologist what the tumor is made of.