Sunday, November 30 | 1:40 p.m.-1:50 p.m. | SSIN01-5 | Room E450A
Johns Hopkins University researchers will share an update on their work to advance AI algorithms for detecting early-stage pancreatic ductal adenocarcinoma (PDAC).
Real early-stage PDAC scans are rare, according to the group led by presenter Wenxuan Li. Therefore, there is a lack of AI training data specific to early-stage cancer.
Envisioning a "Time Machine" PDAC model, Li and colleagues curated 3,144 contrast-enhanced CT scans comprised of 2,098 voxel-wise annotated PDAC cases (divided into tumor stage groups) and 1,046 normal controls. A previously published whole-body segmenter generated masks for the pancreas, vessels, ducts, and 20 other organs in each scan, they noted.
The group employed a reverse-temporal cellular-automata model (conditioned on patient and scan-phase data and tracing tumor growth cues backward), a diffusion model (that blended small, synthetic tumor versions into their original anatomy), and an nnU-Net model (trained on late-stage scans augmented by the synthetic earlier versions).
"A reverse-temporal generative model synthesizes realistic early-stage CT scans from late-stage cases, enabling AI to learn small-tumor patterns without extensive early-stage annotations and improving small tumor (<2 cm) detection sensitivity by 6% over models trained only on real-world tumors," Li and colleagues wrote in their research abstract.
Attend for a discussion of internal validation and external testing performance metrics. Notably, the AI achieved a sensitivity of 36.8% in detecting PDAC at the prediagnostic stage, where radiologists had missed 100% of these PDAC in the original clinical reads of the images.
If validated, the approach offers a scalable strategy to overcome the early-stage data bottleneck and support earlier AI-driven tumor detection, according to the group.
