Related Experiment Video
Updated: Aug 1, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
42.6K
Deep causal learning for pancreatic cancer segmentation in CT sequences.
Chengkang Li1, Yishen Mao2, Shuyu Liang1
1School of Information Science and Technology of Fudan University, Shanghai 200433, China; Key Laboratory of Medical Imaging Computing and Computer Assisted Intervention (MICCAI) of Shanghai, Shanghai 200032, China.
Summary
This study introduces CausegNet, a deep causal learning framework for simultaneously segmenting the pancreas and tumors in 3D CT scans. It significantly improves diagnostic accuracy for pancreatic cancer by reducing background interference.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Pancreatic cancer diagnosis relies on accurate segmentation of the pancreas and tumors in medical images.
- Current deep learning methods struggle with complex backgrounds and low-contrast tissues, leading to segmentation errors.
- Simultaneous segmentation of pancreas and tumor is crucial but challenging due to anatomical variations and tumor obscurity.
Purpose of the Study:
- To develop a novel deep causal learning framework (CausegNet) for simultaneous pancreas and tumor co-segmentation in 3D CT sequences.
- To enhance the understanding of anatomical causal relationships between foreground (pancreas, tumor) and background tissues.
- To improve the robustness and accuracy of segmentation for challenging cases, including deformable pancreases and inconspicuous tumors.
Main Methods:
- Proposed a deep causal learning framework, CausegNet, incorporating a causality-aware module and counterfactual loss.
- Integrated causal inference to differentiate intrinsic foreground features from background interference.
- Developed a sequential search strategy for pancreas and tumor segmentation based on extracted causal features.
Main Results:
- Achieved superior co-segmentation performance on public and clinical datasets, with highest pancreas/tumor Dice coefficients of 86.67%/84.28%.
- Demonstrated causal interpretability and stability through visualized features and anti-noise experiments.
- Improved accuracy and sensitivity in downstream pancreatic cancer risk assessment by 12.50% and 50.00% respectively, outperforming experienced clinicians.
Conclusions:
- CausegNet effectively handles complex backgrounds and anatomical variations for accurate pancreas and tumor co-segmentation.
- The deep causal learning approach enhances segmentation robustness and interpretability.
- This method shows significant potential for improving the accuracy and sensitivity of clinical pancreatic cancer diagnosis and risk assessment.
Keywords:
Counterfactual inferenceDeep causal learningPancreatic cancer co-segmentationPancreatic cancer risk assessment
