Enhancing Uncertainty Estimation in Semantic Segmentation via Monte-Carlo Frequency Dropout

TL;DR

The results show that MC-Frequency Dropout improves calibration, convergence, and semantic uncertainty, thereby improving prediction scrutiny, boundary delineation, and has the potential to enhance medical decision-making.

Abstract

Monte-Carlo (MC) Dropout provides a practical solution for estimating predictive distributions in deterministic neural networks. Traditional dropout, applied within the signal space, may fail to account for frequency-related noise common in medical imaging, leading to biased predictive estimates. A novel approach extends Dropout to the frequency domain, allowing stochastic attenuation of signal frequencies during inference. This creates diverse global textural variations in feature maps while preserving structural integrity -- a factor we hypothesize and empirically show is contributing to accurately estimating uncertainties in semantic segmentation. We evaluated traditional MC-Dropout and the MC-frequency Dropout in three segmentation tasks involving different imaging modalities: (i) prostate zones in biparametric MRI, (ii) liver tumors in contrast-enhanced CT, and (iii) lungs in chest X-ray scans. Our results show that MC-Frequency Dropout improves calibration, convergence, and semantic uncertainty, thereby improving prediction scrutiny, boundary delineation, and has the potential to enhance medical decision-making.

Figures (4)

• Figure 1: Impact of Signal and Frequency Dropout on Structural Integrity. Left subplots show feature maps with traditional (Signal) Dropout ($p$ = 0.1, top) and Frequency Dropout ($p$ = 0.1, bottom). The middle subplots display Sobel gradient maps, while the right subplots illustrate structural changes by comparing pre- and post-Dropout gradients (red indicates intensified gradients, suggesting the emergence of new structures, and blue reflects diminished gradients, revealing blurring and disruption of existing structures). Signal Dropout introduces non-uniform changes in the gradient map, creating new edges while disrupting existing ones. In contrast, Frequency Dropout maintains more uniform changes across imaging features, preserving spatial dependencies. For example, gradient changes follow edges, rather than appearing irregular at the voxel level.
• Figure 2: Calibration of uncertainty estimates for the best dropout configurations (rates and layer placements) of frequency (red) and signal dropout (blue). Subplots show the discrepancy between MC-dropout uncertainties and full (no-dropout) model errors. Frequency dropout aligns uncertainty estimates more accurately across tasks, while preserving segmentation performance similar to the full model (Table \ref{['table1']}).
• Figure 3: Calibration of MC-dropout uncertainty estimates with model errors, measured by Expected Calibration Error (ECE) ($\downarrow$) for different dropout configurations, including dropout rates and layer placements: Encoder Dilution, Decoder Dilution, and Global Dilution.
• Figure 4: MC-dropout uncertainty maps for segmentation tasks using Signal and Frequency Dropout. Large subplots show regions of ground truth (yellow line), full model prediction (blue line), and segmentation errors (red region). Small subplots present uncertainty maps: top for Signal Dropout, bottom for Frequency Dropout, with uncertainty measured as the standard deviation of voxel-level predictions across MC repetitions. Each plot uses the optimal dropout configuration for its approach.