Biologically explicable multimodal model predicting local tumor progression and tumor invasiveness of hepatocellular carcinoma

Purpose: Local tumor progression (LTP) of hepatocellular carcinoma (HCC) after thermal ablation (TA) is related to tumor invasiveness and threaten the health. We aim to build a multimodal model to explicable tumor invasiveness and reduce LTP. Methods: Contrast-enhanced ultrasound (CEUS), magnetic resonance imaging (MRI), biological, clinical and prognostic information were collected to build the model. Long short-term memory network and radiomics were used to extract image features. Logistic regression was used to combined image and clinical information. Pathological, immunohistochemical and RNA sequencing analyses were used to explicable tumor invasiveness. Moderation analysis was used to provide suitable minimum ablation margin (MAM) for high-invasiveness tumors in safe location (not adjacent to vessel or capsule) to reduce LTP. Results: 1208 HCCs were collected as training (n=502), validation (n=180), internal test (n=250) and external test (n=276) sets. AUC of model was 0.809 and 0.811 in internal and external test sets. High-invasiveness group showed higher microvascular invasion proportion, higher macrotrabecular-massive HCC proportion, lower differentiation, higher CK-7 and GPC-3 positive rate and higher VEGFA, MMP-9, HSPA1A expression ( p <0.05). KEGG and GSEA analysis revealed the upregulation of pathways related to angiogenesis, tolerance to stress response, and tumor metastasis in high-invasiveness group. The 8 mm MAM ablation strategy can effectively decrease the LTP incidence of high-invasiveness group (from 42.4% to 10.5%, p =0.027) to the level comparable to low-invasiveness group (10.5% vs. 6.1%, p =0.613) in external test set. Conclusion: Multimodal model achieved satisfactory performance on classifying tumor invasiveness, and provided effective strategy for high-invasiveness tumor to reduce LTP occurrence.