Related Experiment Video
Updated: Nov 2, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.1K
Medical Image Segmentation With Deep Atlas Prior.
IEEE Transactions on Medical Imaging
|June 15, 2021
Summary
This study introduces a novel deep learning framework for medical image segmentation, incorporating anatomical prior knowledge via a Deep Atlas Prior (DAP) loss function. This method enhances organ segmentation accuracy, particularly with limited data.
Area of Science:
- Medical Imaging Analysis
- Artificial Intelligence in Healthcare
- Computational Anatomy
Background:
- Accurate organ segmentation is crucial for computer-aided diagnosis but is hindered by limited annotated data and image quality issues.
- Medical images possess inherent anatomical prior knowledge (organ shape, position) that can significantly improve segmentation accuracy.
- Existing deep learning models often struggle to effectively leverage this anatomical information.
Purpose of the Study:
- To propose a novel deep learning segmentation framework that integrates anatomical prior knowledge directly into the loss function.
- To introduce the Deep Atlas Prior (DAP) loss, which encodes probabilistic atlas information regarding organ location and shape.
- To develop an adaptive Bayesian loss that dynamically balances DAP loss with conventional likelihood losses for improved learning.
Main Methods:
- Developed a Deep Atlas Prior (DAP) loss function utilizing probabilistic atlases for organ location and shape information.
- Integrated DAP loss with Dice and focal losses within a Bayesian framework to create an adaptive Bayesian loss.
- Combined the adaptive Bayesian loss with state-of-the-art fully-supervised and semi-supervised deep segmentation models.
Main Results:
- The proposed framework demonstrated significant improvements in organ segmentation accuracy when integrated with existing deep learning models.
- Validation on public (ISBI LiTS 2017) and private datasets showed enhanced performance for liver and spleen segmentation.
- The adaptive Bayesian loss effectively adjusted the balance between prior and likelihood information during training.
Conclusions:
- The proposed Deep Atlas Prior (DAP) loss integrated into an adaptive Bayesian framework offers a powerful method for enhancing medical image organ segmentation.
- This approach effectively leverages anatomical prior knowledge, overcoming limitations of data scarcity and image quality.
- The framework is versatile and can be applied to various deep segmentation models to boost their performance in clinical applications.
