Contrast-enhanced magnetic resonance imaging radiomics plus deep learning predicts early hepatocellular carcinoma response to drug-eluting bead chemoembolization

A combined magnetic resonance imaging radiomics, deep learning, and clinical model predicted short-term efficacy of drug-eluting bead chemoembolization in hepatocellular carcinoma, with external test AUC 0.804 and internal test AUC 0.822 in a multicenter retrospective study.

• The model used contrast-enhanced MRI arterial phase, diffusion-weighted imaging, and T2-weighted imaging; arterial-phase deep learning features were extracted with a pretrained 3D ResNet-18.
• Microsphere type (p=0.042) and combined arterial/diffusion radiomics-deep learning scores (p=0.01) were independently associated with short-term efficacy.
• A single arterial-phase radiomics model had a training AUC of 0.900 but lower internal test AUC (0.750), while the combined AP+DWI model generalized better.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.