MRI-based automatic identification and segmentation of extrahepatic cholangiocarcinoma using deep learning network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate identification and segmentation of extrahepatic cholangiocarcinoma (ECC) from MRI is challenging due to small tumor size and complex anatomy.
• Manual delineation is time-consuming and limited, necessitating automated methods for ECC identification and segmentation.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automatic identification and segmentation of ECC using MRI.
• To compare the performance of single-mode (T1WI, T2WI, DWI) and combined-mode deep learning models for ECC detection.

Main Methods:

• A dataset of 137 ECC patients (C1) and 40 external validation patients (C2) underwent T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and diffusion-weighted imaging (DWI).
• A 3D VB-Net deep learning model was trained to create single-mode identification and segmentation models for T1WI, T2WI, and DWI.
• Model performance was assessed using training, testing, and external validation sets, with manual delineations serving as ground truth.

Main Results:

• The diffusion-weighted imaging (DWI)-based model (Model 3) demonstrated superior identification performance across all cohorts (success rates: 0.980 training, 0.786 testing, 0.725 external validation).
• Model 3 achieved high Dice similarity coefficients (DSC) for ECC segmentation: 0.922 (training), 0.495 (testing), and 0.466 (external validation).

Conclusions:

• The DWI-based deep learning model significantly outperformed T1WI and T2WI models in automatically identifying and segmenting extrahepatic cholangiocarcinoma.
• This DWI-based approach holds potential for improving clinical decisions and aiding in prognosis determination for ECC patients.