Related Experiment Video
Updated: Sep 24, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
EVALUATION OF COMPLEXITY MEASURES FOR DEEP LEARNING GENERALIZATION IN MEDICAL IMAGE ANALYSIS
Aleksandar Vakanski1, Min Xian2
1Department of Nuclear Engineering and Industrial Management, University of Idaho, Idaho Falls, USA.
IEEE International Workshop on Machine Learning for Signal Processing : [Proceedings]. IEEE International Workshop on Machine Learning for Signal Processing
|May 9, 2022
Summary
Deep learning models for medical imaging struggle with generalization across different data sources. PAC-Bayes flatness and path norm measures best predict generalization for breast ultrasound classifiers.
Area of Science:
- Medical image analysis
- Deep learning
- Generalization capacity
Background:
- Deep learning models in medical imaging often exhibit poor generalization on data from different sources (e.g., devices, patient populations).
- Accurate assessment of generalization is vital for clinical trust, but current complexity measures and generalization bounds show discrepancies with actual performance.
- Existing empirical studies often use general-purpose image datasets, limiting their applicability to specific medical domains.
Purpose of the Study:
- To empirically investigate the correlation between 25 complexity measures and the generalization abilities of deep learning classifiers for breast ultrasound images.
- To identify which complexity measures are most effective in predicting generalization performance in this specific medical context.
- To evaluate the impact of a multi-task approach (classification and segmentation) on generalization.
Main Methods:
- An empirical study was conducted using deep learning classifiers trained on breast ultrasound images.
- 25 different complexity measures were calculated for the trained models.
- The correlation between these complexity measures and the models' generalization performance on unseen data was analyzed.
Main Results:
- PAC-Bayes flatness and path norm measures demonstrated the most consistent correlation with generalization performance across different model-data combinations.
- These specific complexity measures offer a more reliable prediction of generalization compared to others evaluated.
- Employing a multi-task learning approach, combining classification and segmentation, was found to enhance generalization capabilities.
Conclusions:
- PAC-Bayes flatness and path norm are promising metrics for assessing and predicting the generalization of deep learning models in breast ultrasound analysis.
- The findings suggest that these measures can help improve the trustworthiness of AI in medical imaging by providing better estimates of real-world performance.
- Multi-task learning strategies, integrating segmentation with classification, can be effectively leveraged to improve the generalization of deep learning models for breast ultrasound applications.