What Can Conformal Risk Control Certify for PET/CT Tumour Segmentation?

Weiyue Zheng, Surajit Ray
Proceedings of the Fifteenth Symposium on Conformal and Probabilistic Prediction with Applications, PMLR 329:1063-1065, 2026.

Abstract

A conformal guarantee for segmentation is useful only when the controlled loss reflects the clinical decision. In PET/CT tumour imaging, missing a whole lesion can matter more than small boundary errors. We apply connected-component risk-controlling prediction sets (RCPS) to fixed, pretrained LesionTracer probability maps for 900 autoPET cases, asking whether one global threshold can certify a binary missed-lesion loss and a continuous voxel-level missed-tumour loss. At $\alpha$ = $\delta$ = 0.1, the binary loss cannot be certified: on lesion-positive calibration cases its empirical risk is 0.177 even at the smallest threshold we evaluated. The voxel loss behaves differently: an empirical-Bernstein bound certifies $\hat{\lambda} = 0.78$, giving P(Rvox($\hat{\lambda}$) $\leq$ 0.1) $\geq$ 0.9, once lesion-free cases contribute zero loss. This binary infeasibility holds for coverage requirements $\gamma$ from 0.5 to 0.9. We read the contrast as a statement about the prediction family: a global threshold cannot recover lesions the network never supported.

Cite this Paper

BibTeX
@InProceedings{pmlr-v329-zheng26a, title = {What Can Conformal Risk Control Certify for PET/CT Tumour Segmentation?}, author = {Zheng, Weiyue and Ray, Surajit}, booktitle = {Proceedings of the Fifteenth Symposium on Conformal and Probabilistic Prediction with Applications}, pages = {1063--1065}, year = {2026}, editor = {Ahlberg, Ernst and Johansson, Ulf and Boström, Henrik and Carlevaro, Alberto and Hallberg Szabadváry, Johan and Carlsson, Lars}, volume = {329}, series = {Proceedings of Machine Learning Research}, month = {02--04 Sep}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v329/main/assets/zheng26a/zheng26a.pdf}, url = {https://proceedings.mlr.press/v329/zheng26a.html}, abstract = {A conformal guarantee for segmentation is useful only when the controlled loss reflects the clinical decision. In PET/CT tumour imaging, missing a whole lesion can matter more than small boundary errors. We apply connected-component risk-controlling prediction sets (RCPS) to fixed, pretrained LesionTracer probability maps for 900 autoPET cases, asking whether one global threshold can certify a binary missed-lesion loss and a continuous voxel-level missed-tumour loss. At $\alpha$ = $\delta$ = 0.1, the binary loss cannot be certified: on lesion-positive calibration cases its empirical risk is 0.177 even at the smallest threshold we evaluated. The voxel loss behaves differently: an empirical-Bernstein bound certifies $\hat{\lambda} = 0.78$, giving P(Rvox($\hat{\lambda}$) $\leq$ 0.1) $\geq$ 0.9, once lesion-free cases contribute zero loss. This binary infeasibility holds for coverage requirements $\gamma$ from 0.5 to 0.9. We read the contrast as a statement about the prediction family: a global threshold cannot recover lesions the network never supported.} }
Endnote
%0 Conference Paper %T What Can Conformal Risk Control Certify for PET/CT Tumour Segmentation? %A Weiyue Zheng %A Surajit Ray %B Proceedings of the Fifteenth Symposium on Conformal and Probabilistic Prediction with Applications %C Proceedings of Machine Learning Research %D 2026 %E Ernst Ahlberg %E Ulf Johansson %E Henrik Boström %E Alberto Carlevaro %E Johan Hallberg Szabadváry %E Lars Carlsson %F pmlr-v329-zheng26a %I PMLR %P 1063--1065 %U https://proceedings.mlr.press/v329/zheng26a.html %V 329 %X A conformal guarantee for segmentation is useful only when the controlled loss reflects the clinical decision. In PET/CT tumour imaging, missing a whole lesion can matter more than small boundary errors. We apply connected-component risk-controlling prediction sets (RCPS) to fixed, pretrained LesionTracer probability maps for 900 autoPET cases, asking whether one global threshold can certify a binary missed-lesion loss and a continuous voxel-level missed-tumour loss. At $\alpha$ = $\delta$ = 0.1, the binary loss cannot be certified: on lesion-positive calibration cases its empirical risk is 0.177 even at the smallest threshold we evaluated. The voxel loss behaves differently: an empirical-Bernstein bound certifies $\hat{\lambda} = 0.78$, giving P(Rvox($\hat{\lambda}$) $\leq$ 0.1) $\geq$ 0.9, once lesion-free cases contribute zero loss. This binary infeasibility holds for coverage requirements $\gamma$ from 0.5 to 0.9. We read the contrast as a statement about the prediction family: a global threshold cannot recover lesions the network never supported.
APA
Zheng, W. & Ray, S.. (2026). What Can Conformal Risk Control Certify for PET/CT Tumour Segmentation?. Proceedings of the Fifteenth Symposium on Conformal and Probabilistic Prediction with Applications, in Proceedings of Machine Learning Research 329:1063-1065 Available from https://proceedings.mlr.press/v329/zheng26a.html.

Related Material