August 24, 2026
PAC-AID Talks: Sovanlal Mukherjee on catching pancreatic cancer before it can be seen
The next PAC-AID Talk is Friday, September 4, 2026 at 9:00 AM HST on Zoom. Sovanlal Mukherjee, PhD, Assistant Professor of Radiology at Mayo Clinic in Rochester, Minnesota, presents “Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability.”
Why this talk matters
Pancreatic ductal adenocarcinoma (PDA) is among the deadliest cancers, and the reason is timing: more than 85% of cases are found at an unresectable stage. Sporadic PDA — which accounts for 85–90% of all cases — still has no recognised early detection strategy. The obstacle is not carelessness but physics and biology. During the curable pre-clinical window, the pancreas simply looks normal on a CT scan, even to an expert re-reading it later. There is no mass to find.
Dr. Mukherjee’s work asks whether the disease is nonetheless already written into the image, below the threshold of human vision.
REDMOD
REDMOD (Radiomics-based Early Detection MODel) is a fully automated framework that looks for subvisual radiomic signatures of pre-diagnostic PDA on standard-of-care CT. It pairs AI-driven volumetric pancreas segmentation with a heterogeneous ensemble classifier — logistic regression, random forest, and XGBoost combined by soft voting — trained on a 40-feature radiomic signature distilled from 968 candidate features.
Critically, it was validated the hard way. Rather than the balanced case-control design common in earlier work, REDMOD was tested at a ~6:1 control-to-case ratio that approximates what an actual early detection programme would face.
On an independent test set of 493 scans, REDMOD:
- Detected occult PDA with an AUC of 0.82 and 73.0% sensitivity, at a median lead time of 475 days before clinical diagnosis
- Achieved nearly twice the sensitivity of board-certified radiologists (73.0% vs 38.9%), a gap that widened to nearly threefold (68.0% vs 23.0%) for scans taken more than 24 months before diagnosis
- Held 90–92% test–retest concordance across serial scans — the first demonstration that a pre-clinical radiomic signal is longitudinally stable
- Generalised its specificity to independent multi-institutional (81.3%, n=539) and public NIH-PCT (87.5%, n=80) cohorts
There is also a mechanistic finding worth the hour on its own: 90% of the selected signature came from multi-scale wavelet-filtered textural features, which significantly outperformed unfiltered ones (AUC 0.82 vs 0.74). That points to the earliest detectable manifestation of PDA being a subtle, diffuse disruption of pancreatic architecture rather than any discrete lesion — which in turn reframes what an early-detection tool should even be looking for.
About the speaker
Sovanlal Mukherjee, PhD, develops artificial intelligence, machine learning, and quantitative imaging approaches for cancer detection and characterization, with particular emphasis on the early detection of pancreatic cancer. His work spans radiomics, deep learning, predictive modeling, multimodal imaging, radiation therapy planning, and low-dose CT image enhancement, and he is the lead developer of REDMOD.
He is a recipient of the Mayo Clinic Transformative Science Award as part of the Pancreas Cancer Artificial Intelligence Team, as well as the Mayo Clinic Radiology Research Grant Award and a Strategic Pilot Study Grant from the Mayo Clinic Comprehensive Cancer Center.
The paper
The work discussed in this talk is published open access in Gut:
Mukherjee S, Antony A, Patnam NG, et al. Next-generation AI for visually occult pancreatic cancer detection in a low-prevalence setting with longitudinal stability and multi-institutional generalisability. Gut 2026. doi:10.1136/gutjnl-2025-337266
PAC-AID Talks meets monthly, usually the first Friday, at 9:00 AM HST via Zoom. See the events page for the full schedule, and past talks on the YouTube channel. To suggest a speaker, contact abunnell@hawaii.edu.