Abstract
<title>Abstract</title> <p>Background: Tumor-associated macrophages with M2 polarization (TAMs-M2) was a key immunohistochemical phenomenon in hepatocellular carcinoma (HCC), which associated with therapy resistance and poor prognosis. However, diagnosis of TAMs-M2 highly relies on invasive biopsy, which reduced its clinical value. Objective: To develop a multimodal model combining ultrasound (US), magnetic resonance imaging (MRI) and clinical information for TAMs-M2 diagnosis non-invasively and evaluate model value in prognosis. Methods: US image, MRI image, clinical information, immunohistochemical result on TAMs-M2, and prognostic information were retrospectively collected from six centers. Multimodal model based on US deep learning feature, MRI radiomics feature and clinical information was developed to non-invasively predict the TAMs-M2 in HCC. Patients with targeted therapy were classified as high-risk and low-risk groups by multimodal model, and the treatment response and overall survival (OS) were compared in two groups. Results: Since January 2018 and December 2024, 552 eligible HCCs were enrolled and divided into training (n=222), validation (n=111), internal test (n=111), and external test (n=108) sets. Multimodal model demonstrated well diagnostic performance in internal (AUC=0.802, 95%CI: 0.689-0.901) and external (AUC=0.797, 95%CI: =0.689-0.888) test sets. Since January 2020 and December 2024, 225 eligible HCCs were collected from other three centers as prognosis test set, and HCCs in low-risk group showed significantly higher objective response rate (11.2% vs 5.5%, p=0.032), disease control rate (57.3% vs 23.3%, p<0.001), and longer OS (p=0.007) than high-risk group. Conclusion: Multimodal model can accurately and non-invasively predict TAMs-M2 in HCC and effectively stratify the prognosis of patient received targeted therapy, holding promise for guiding personalized treatment strategies.</p>