# Enhancing online adaptive radiotherapy with uncertainty based segmentation error and out-of-distribution detection

## TL;DR

This study explores how uncertainty estimation in deep learning can improve the accuracy of segmenting prostate cancer images for radiotherapy by identifying uncertain predictions and distinguishing between different types of data.

## Contribution

The study introduces uncertainty estimation methods to detect segmentation errors and out-of-distribution data in radiotherapy imaging.

## Key findings

- Uncertainty estimation correlates with segmentation accuracy, as shown by Dice scores for prostate cancer images.
- Predictive entropy effectively separates correct and incorrect segmentations, with higher values at segmentation borders.
- Mutual information achieves 100% separation between in-distribution and out-of-distribution data.

## Abstract

Anatomical segmentation is one of the biggest sources of uncertainty in the online adaptive radiotherapy workflow. The aim of this study was to investigate the relation between the estimated uncertainty in deep learning (DL)-based segmentation and the correctness of the segmentations. In addition, the ability to capture out-of-distribution (OOD) data with uncertainty estimation was tested.

The Monte Carlo dropout method was applied to estimate the uncertainty of a DL model for magnetic resonance (MR)-guided radiotherapy prostate cancer images, trained to segment the clinical target volume (CTV), bladder, and rectum. The training/validation set consisted of 151 T2 MR scans from 26 patients, while the test set consisted of 65 scans from 10 patients. Predictive entropy (PE) was used to capture predictive (model and data) uncertainty. The PE distributions for correct and incorrect predictions were used to find a threshold value. Predicted segmentations with PE values above this threshold value were allocated to the “uncertain group,” and those below to the “certain group.” Dice scores were computed for both groups, using manual segmentations as ground truth. Mutual information (MI) was additionally used to capture epistemic (model) uncertainty as a means to separate in-distribution (ID) from OOD data. Balanced steady-state free precession MRI scans of 10 healthy volunteers were used as OOD data.

The segmentation model obtained Dice scores of 85.7% for the CTV, 94.8% for the bladder, and 86.6% for the rectum. The highest PE values were found at the segmentation borders. Higher PE threshold values resulted in better separation between the certain and uncertain groups. This shows the ability to detect incorrect predictions with uncertainty estimation. A 100% separation between ID and OOD data was achieved with MI.

Uncertainty estimation from a DL-based segmentation model was seen to correlate with Dice scores for segmentation of MR-guided radiotherapy prostate cancer images. This implies that uncertainty estimation could be used to label the quality of the segmentations in the online adaptive radiotherapy workflow. Preliminary results showed that uncertainty estimation could be used to distinguish between ID and OOD data.