Deep Learning Model With Convolutional Neural Network for Detecting and Segmenting Hepatocellular Carcinoma in CT: A...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Hepatocellular carcinoma (HCC) is a prevalent global malignancy.
• Early detection and precise diagnosis are critical for effective patient management.
• Dynamic contrast-enhanced computed tomography (CT) is a key imaging modality for HCC evaluation.

Purpose of the Study:

• To develop and evaluate a U-Net based convolutional neural network (CNN) model for identifying and segmenting HCC lesions.
• To utilize dynamic contrast-enhanced CT images across arterial, venous, and delayed phases.
• To assess the model's performance against expert radiologists.

Main Methods:

• A retrospective study using CT image sets from 110 patients with histopathology-confirmed HCC.
• Implementation of a U-Net architecture CNN with domain adaptation for segmentation.
• Evaluation of the model's sensitivity for HCC identification and Dice score for segmentation accuracy.

Main Results:

• The CNN model demonstrated 100% sensitivity in identifying HCC lesions.
• The median Dice score for HCC segmentation between the CNN model and radiologists was 0.81.
• The model showed high performance in both detection and segmentation tasks.

Conclusions:

• Deep learning models, specifically CNNs, show significant promise for HCC identification and segmentation.
• The developed U-Net based CNN model offers a high-performance tool for analyzing dynamic CT scans in HCC diagnosis.
• AI-assisted analysis can improve the accuracy and efficiency of HCC detection and segmentation.