Related Experiment Video
Updated: Dec 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.2K
Automatic Pancreas Segmentation in CT Images With Distance-Based Saliency-Aware DenseASPP Network
IEEE Journal of Biomedical and Health Informatics
|September 11, 2020
Summary
This study introduces DSD-ASPP-Net, a novel deep learning model for accurate pancreas segmentation. The model enhances segmentation by integrating coarse predictions with saliency maps, improving diagnostic capabilities for pancreas diseases.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Pancreas segmentation is crucial for diagnosing and managing pancreatic diseases.
- Deep neural networks face challenges in segmenting small, low-contrast organs like the pancreas due to their flexible anatomy.
- Existing coarse-to-fine methods for pancreas segmentation neglect rich image context by only using object location.
Purpose of the Study:
- To propose a novel distance-based saliency-aware model, DSD-ASPP-Net, for improved pancreas segmentation.
- To leverage coarse segmentation predictions more effectively by highlighting pancreas features and incorporating image context.
- To enhance the accuracy of fine-stage segmentation for challenging pancreatic imaging.
Main Methods:
- A Dense Atrous Spatial Pyramid Pooling (DenseASPP) model was trained to generate pancreas location and probability maps.
- Geodesic distance-based saliency transformation was applied to convert probability maps into saliency maps.
- Saliency-aware modules were integrated into DenseASPP in the fine stage, combining saliency maps with image context to create DSD-ASPP-Net.
Main Results:
- The DSD-ASPP-Net architecture utilizes multi-scale feature representation and a larger receptive field for dense feature extraction.
- The model effectively addresses challenges posed by variable object sizes and locations in pancreas segmentation.
- An average Dice-Sørensen Coefficient (DSC) of 85.49±4.77% was achieved on the NIH pancreas dataset, outperforming previous coarse-to-fine methods.
Conclusions:
- The proposed DSD-ASPP-Net significantly improves pancreas segmentation accuracy by effectively utilizing coarse segmentation information and image context.
- The saliency-aware approach enhances feature highlighting, leading to more robust segmentation of the pancreas.
- This method offers a promising advancement for computer-aided diagnosis and prognosis of pancreas-related conditions.
