Related Experiment Video
Updated: Jul 7, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Anatomically constrained deep learning for clinical-grade volumetric pancreatic cancer segmentation: development,
Sovanlal Mukherjee1, Khurram Khaliq Bhinder1, Armin Zarrintan1
1Department of Radiology, Mayo Clinic, Rochester, MN, USA.
NPJ Precision Oncology
|July 3, 2026
Summary
A new AI model, Model-BB, accurately segments pancreatic ductal adenocarcinoma (PDAC) on CT scans. This automated segmentation is crucial for precision oncology, showing robust performance across diverse datasets and outperforming other advanced models.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Automated segmentation of pancreatic ductal adenocarcinoma (PDAC) is essential for precision oncology.
- Current segmentation methods lack sufficient validation for clinical deployment due to data heterogeneity and scale.
Purpose of the Study:
- To develop and validate a robust AI model for automated PDAC segmentation on CT scans.
- To assess the model's performance across multi-institutional datasets and diverse acquisition parameters.
Main Methods:
- Developed and validated Model-BB, a 3D convolutional neural network, using 1859 multi-institutional, treatment-naïve PDAC CT scans.
- Evaluated model performance using Dice Similarity Coefficient (DSC) on internal and external validation sets.
- Compared Model-BB against Swin UNETR and human reader performance on challenging cases.
Main Results:
- Model-BB achieved a DSC of 0.76 ± 0.13 (internal) and 0.76 ± 0.09 (external validation).
- Performance remained stable across different acquisition sites, scanner vendors, slice thicknesses, and time periods.
- Model-BB outperformed Swin UNETR (DSC 0.76 vs. 0.68, p < 0.001) and exceeded individual reader-pair agreement on difficult cases.
Conclusions:
- Anatomically constrained, task-specific segmentation using Model-BB provides a reproducible geometric foundation for quantitative tumor burden assessment.
- The model supports treatment response evaluation and multimodal outcomes modeling.
- Prospective validation in clinical trial workflows is recommended.
