Related Experiment Video
Updated: Oct 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
PAC Bayesian Performance Guarantees for Deep (Stochastic) Networks in Medical Imaging
Anthony Sicilia1, Xingchen Zhao2, Anastasia Sosnovskikh2
1Intelligent Systems Program, University of Pittsburgh, Pittsburgh, USA.
Medical Image Computing and Computer-Assisted Intervention : MICCAI ... International Conference on Medical Image Computing and Computer-Assisted Intervention
|December 27, 2021
Summary
This study applies PAC-Bayesian methods to medical imaging, offering mathematical bounds on deep learning generalization. These bounds improve explainability and reduce the need for holdout sets in small datasets.
Area of Science:
- Medical Imaging
- Machine Learning
- Computer Vision
Background:
- Deep neural networks (DNNs) are widely used in medical imaging.
- A key challenge is DNNs' tendency to overfit small datasets, common in medical imaging.
- Overfitting limits the generalization ability of DNN models.
Purpose of the Study:
- To apply PAC-Bayesian framework for generalization error bounds in DNNs for medical imaging.
- To address overfitting concerns in small medical imaging datasets.
- To evaluate these bounds on classification and segmentation tasks.
Main Methods:
- Utilized the PAC-Bayesian framework to derive generalization error bounds.
- Applied techniques to the ISIC 2018 challenge dataset, a small medical imaging dataset.
- Evaluated bounds for both classification and segmentation tasks.
Main Results:
- The derived PAC-Bayesian bounds were competitive with simpler baseline methods.
- The bounds provided enhanced explainability for network performance.
- The approach reduced the necessity of using holdout sets for validation.
Conclusions:
- PAC-Bayesian bounds offer a viable method to quantify generalization in DNNs for medical imaging.
- This framework enhances trust in DNN models trained on limited medical data.
- The method is effective for both classification and segmentation, improving model reliability.
