Semi-automatic fine delineation scheme for pancreatic cancer

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Pancreatic cancer delineation in medical images presents challenges due to large data volumes and patient variability.
• Accurate segmentation of pancreatic tumors is crucial for effective treatment planning.

Purpose of the Study:

• To develop a semi-automatic scheme for precise and rapid delineation of pancreatic cancer in computed tomography (CT) images.
• To reduce the workload of physicians involved in medical image analysis.

Main Methods:

• A 3D Res U-Net deep learning model was employed for segmenting cancer in image blocks.
• Physicians manually delineated the start and end slices, with the model predicting cancer in intermediate slices.
• The study evaluated model performance and physician workload across various image block sizes.

Main Results:

• The optimal image block size of 5 minimized physician workload while ensuring accurate cancer delineation.
• Achieved a Dice similarity coefficient of 0.894 ± 0.029 and a 95% Hausdorff distance of 3.465 ± 0.710 mm.
• The semi-automatic method demonstrated a 2.16 times increase in speed compared to manual delineation.

Conclusions:

• The proposed 3D semi-automatic method effectively delineates pancreatic cancer in CT images, addressing challenges like class imbalance and non-rigid features.
• This approach significantly reduces physician workload and is expected to enhance clinical efficiency.
• The block prediction strategy aids in accurate segmentation and improved diagnostic workflows.