Related Experiment Video
Updated: Jul 22, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
Trustworthy Medical Image Segmentation with improved performance for in-distribution samples.
Sneha Shukla1, Lokendra Birla2, Anup Kumar Gupta1
1Indian Institute of Technology Indore, Indore, India.
Summary
This study introduces TrustMIS, a new method to improve the trustworthiness of Deep Learning models in medical image segmentation. TrustMIS enhances predictions from in-distribution samples, addressing critical issues in medical AI.
Area of Science:
- Artificial Intelligence
- Medical Imaging
- Computer Vision
Background:
- Deep Learning (DL) models achieve high performance in Medical Image Segmentation (MIS) but suffer from non-transparent predictions.
- Erroneous In-Distribution (ID) and Out-Of-Distribution (OOD) samples lead to distrust and potential harm in medical applications.
- Existing methods focus on OOD detection, but trustworthiness of ID samples in MIS remains a significant challenge.
Purpose of the Study:
- To propose a novel method, TrustMIS (Trustworthy Medical Image Segmentation), to enhance the trustworthiness and performance of DL-based MIS models for ID samples.
- To address the critical need for reliable predictions in medical image analysis.
- To provide a quantitative measure of trustworthiness for MIS models.
Main Methods:
- TrustMIS employs a three-fold approach: Investigating Trustworthiness (IT), Improving Non-Trustworthy prediction (INT), and Classifier Switching Operation (CSO).
- The IT method analyzes input characteristics and consistency to assess trustworthiness.
- The INT method refines segmentation by leveraging input transformations (e.g., rotation), and CSO selects the most trustworthy model predictions.
Main Results:
- TrustMIS successfully provides a trustworthiness measure for DL-based MIS models.
- The method demonstrates improved performance on state-of-the-art MIS models using publicly available datasets.
- Experimental results show TrustMIS outperforms existing methods in enhancing trustworthiness and segmentation accuracy.
Conclusions:
- TrustMIS offers a robust solution for improving the reliability of Deep Learning models in Medical Image Segmentation.
- The proposed method enhances the trustworthiness of predictions from in-distribution samples, a crucial step for clinical adoption.
- The availability of the implementation facilitates further research and application of trustworthy AI in medical imaging.
