News

Best Pan-cancer Segmentation Accuracy at FLARE 2026

October 2026

FLARE 2026 Best Pan-cancer Segmentation Accuracy Winner Award certificate, MICCAI 2026

Our pan-cancer lesion segmentation model won the Best Pan-cancer Segmentation Accuracy award at FLARE 2026, the Fast, Low-resource, Accurate, Robust and Effectual medical image analysis challenge held at MICCAI 2026.

Measuring tumors on CT is still largely manual. A radiologist picks a few target lesions, measures them by hand and tracks how they change over time. We built a single model that finds and outlines every lesion, across all organs, from one CT scan.

Two things made it hard. Lesion volumes span four orders of magnitude, from sub-millilitre nodules to litre-scale masses, and more than half of the lesions we trained on were under 1 mL. And we combined 28 public datasets, about 17,000 CT scans, annotated under different protocols, most of which label only the target lesion.

On the challenge's hidden validation set the model reached 81.2% Dice and 76.4% normalized surface Dice, running in about 18 seconds per scan with around 2.5 GB of peak GPU memory.

The paper, Pan-cancer Lesion Segmentation in CT under Heterogeneous and Partial Annotation, is by Luca Pegolotti, Virginia Tasso, Marco Santambrogio and Angelica Iacovelli. The work is a collaboration between Nucleo and NECSTLab at Politecnico di Milano. The full paper will follow.

Sources

All news