Deep learning with attention modules and residual transformations improves hepatocellular carcinoma (HCC)...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Hepatocellular carcinoma (HCC) detection remains a challenge.
• Conventional deep learning models may not fully capture complex imaging features.
• There is a need for improved accuracy in differentiating HCC from other liver lesions.

Purpose of the Study:

• To develop and evaluate a novel deep learning model (GAN+SA+ResNeXt) for enhanced HCC prediction.
• To investigate the efficacy of combining generative adversarial networks (GAN) with self-attention (SA) and ResNeXt for HCC differentiation.
• To improve the accuracy and efficiency of HCC diagnosis using multiphase CT scans.

Main Methods:

• Retrospective analysis of 228 multiphase CT scans from 57 patients (30 HCC, 27 non-HCC).
• Implementation of automatic liver segmentation and Hounsfield unit (HU) normalization.
• Training and evaluation of 3D GAN, 3D GAN+A, and 3D GAN+A+ResNeXt models using five-fold cross-validation.

Main Results:

• The GAN+SA+ResNeXt model achieved an Area Under the Receiver Operating Characteristics Curve (AUROC) of 95%.
• The proposed model demonstrated high accuracy (91%), sensitivity (93%), and acceptable specificity (88%).
• The prediction time was notably efficient at 0.04s.

Conclusions:

• The proposed GAN+SA+ResNeXt deep learning model is feasible and effective for HCC diagnosis from CT.
• This approach offers improved accuracy and efficiency in differentiating HCC from other liver lesions.
• The model shows significant clinical potential for improving liver lesion characterization.