Reproduce “Report Supervision” on OpenResearch

About this paper

Pedro R. A. S. Bassia, Wenxuan Li, Jakob Wasserthal, Jieneng Chen, Xinze Zhou, Zheren Zhu, Chuntung Zhuanga, Sergio Decherchi and 5 more

R-Super turns routine radiology reports into direct supervision for 3D tumor segmentation, using reported lesion volume, count, size, and anatomical location to constrain CT predictions without requiring voxel-level masks. Across pancreatic and kidney tumor datasets, it improves detection and segmentation under limited annotation, including a rise in external pancreatic F1-score from 62% to 67% with only 50 masks and from 80% to 85% on a large cross-institutional evaluation; it also improves performance on non-contrast CT.

Key insights

• Reports can supervise segmentation directly when their qualitative descriptions are converted into differentiable constraints on the output mask, rather than being used only for auxiliary classification or representation learning.
• Volume information provides a stable optimization signal, while the Ball Loss adds lesion-level localization and multiplicity; combining both is generally more effective than using either alone.
• Report-derived size and anatomical sub-segment information materially affect performance, yet the method remains reasonably tolerant of annotation-extraction errors and can exploit reports from imaging protocols absent from the mask dataset.

Summary by alphaXiv.