Related Experiment Video
Updated: Nov 18, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.1K
Unraveling the deep learning gearbox in optical coherence tomography image segmentation towards explainable
Peter M Maloca1,2,3,4, Philipp L Müller5,6, Aaron Y Lee7,8,9
1Institute of Molecular and Clinical Ophthalmology Basel (IOB), Basel, Switzerland. peter.maloca@iob.ch.
Communications Biology
|February 6, 2021
Summary
This study introduces Traceable Relevance Explainability (T-REX) to make convolutional neural networks for optical coherence tomography image segmentation more transparent. The T-REX technique achieved high accuracy, comparable to human graders, enhancing machine learning interpretability in medical imaging.
Area of Science:
- Medical imaging analysis
- Artificial intelligence in healthcare
- Computer vision
Background:
- Machine learning (ML) models excel at medical data analysis but often lack transparency.
- Understanding the internal decision-making processes of ML algorithms is crucial for clinical adoption.
- Optical coherence tomography (OCT) is vital for visualizing subsurface tissue structures.
Purpose of the Study:
- To enhance the transparency of a convolutional neural network (CNN) used for OCT image segmentation.
- To develop and evaluate a novel explainability technique, Traceable Relevance Explainability (T-REX), for ML models in medical imaging.
- To quantify the performance of the ML algorithm against human expert annotations.
Main Methods:
- Development of a T-REX technique integrating ground truth generation by multiple graders.
- Calculation of Hamming distances to compare ML algorithm predictions with human grader variability.
- Implementation of a 'neural recording' smart data visualization for ML process comprehension.
- Application of a CNN for OCT image segmentation.
Main Results:
- The ML algorithm demonstrated an average variability of 1.75% compared to human graders, which was slightly lower than the 2.02% variability among human graders.
- Ambiguity in ground truth data was shown to significantly impact ML results, and this impact was visualized.
- The CNN exhibited balanced performance across different graders and allowed for adaptable predictions based on specific image compartments.
- The T-REX setup successfully rendered the ML segmentation process more transparent and understandable.
Conclusions:
- The T-REX technique provides a valuable method for increasing the transparency and interpretability of CNNs in medical image analysis.
- The findings highlight the impact of ground truth quality on ML model performance and the utility of explainability tools in identifying such issues.
- This approach facilitates a better understanding of ML decision-making, paving the way for optimized and trustworthy AI applications in healthcare.
- The T-REX system enables ML models to achieve performance comparable to human experts while offering insights into their operational logic.
