COMPASS: Robust Feature Conformal Prediction for Medical Segmentation Metrics

Uncertainty quantification is of critical need in medical image analysis, a field used for decision support in high-stakes clinical diagnosis and treatment planning applications [1, 2]. A fundamental task in medical image analysis is image segmentation, the task of separating anatomical structures and lesions from each other within an image. Deep learning models, particularly U-Net variants [3, 4], have achieved state-of-the-art performance in medical image segmentation. In practice, the outputs of these models ("segmentation maps") are often treated as an intermediate result that are then used to automatically derive downstream metrics of interest (known as "radiomics"), such as the areas/volumes or texture patterns of specific anatomic regions (Figure 1, left). These derived metrics are then used for decision support to guide clinicians in diagnosis and treatment. Conformal prediction (CP) has emerged as a popular, statistically principled uncertainty quantification framework of choice in machine learning, providing guarantees without restrictive distributional assumptions [5, 6, 7, 8, 9]. While well-studied in the context of typical prediction tasks involving scalar output variables, CP is less explored for tasks such as medical image segmentation, in which the output variables are images. Existing CP methods for segmentation typically focus on deriving bounds for pixel-level errors [10, 11, 12, 13], which, while useful for understanding variations of local segmentation contours, may yield meaningless or misaligned intervals for downstream derived metrics. On the other hand, a recent study shows that treating the segmentation-to-metric pipeline as a black box and performing CP directly on the output metric space yields intervals that are well-aligned to the metrics (by construction), but are also often inefficient (i.e., large) because the internal biases of the pipeline are not exploited in the vanilla CP formulation [14]. 1