Uncertainty-Aware Brain-Tumor Type Classification From MRI: Segmentation-Guided ROI, Conformal Prediction, and Selective Abstention Under Distribution Shift
TL;DR
An uncertainty-aware classification framework is presented that couples segmentation-guided region-of-interest (ROI) localization with post-hoc calibration, split-conformal prediction, and selective abstention, and a deferral rule that abstains on unreliable inputs.
Abstract
Reliable classification of brain-tumor type from magnetic resonance imaging (MRI) is essential for treatment planning and patient management. While deep neural networks achieve high accuracy on this task, they typically produce point predictions without a measure of reliability, and their confidence need not fall when image quality departs from the training distribution. An uncertainty-aware classification framework is presented that couples segmentation-guided region-of-interest (ROI) localization with post-hoc calibration, split-conformal prediction, and selective abstention: a U-Net localizes the tumor; the cropped region is classified; and the prediction is equipped with a calibrated confidence, a conformal set with a distribution-free coverage guarantee, and a deferral rule that abstains on unreliable inputs. All experiments used the Figshare brain-tumor dataset under leakage-free patient-level cross-validation. ROI localization improved patient-level balanced accuracy from 93.8% to 96.1% with a predicted mask while also improving calibration, and the 90% conformal coverage held under deployment with a predicted mask. Under bias-field, blur, and gamma corruptions, where exchangeability no longer holds, the empirical coverage degraded but was restored by recalibration and patient-aware inference, while the abstention rate rose with severity. The exchangeable unit for a valid guarantee is shown to be the patient, not the slice: slice-level calibration met the intended per-patient coverage in only 27% of resampled splits, versus 84% for patient-level calibration. The reliability layer further transfers to a four-class benchmark that adds healthy scans.