LLM-Enhanced Multimodal Fusion of SPECT Radiomics and Clinical Data for Predicting 131I Therapeutic Response in Differentiated Thyroid Cancer.

BACKGROUND

Patients with differentiated thyroid cancer (DTC) treated with radioactive iodine (RAI) remain at risk of adverse outcomes. Early prediction of RAI efficacy is critical for optimizing clinical management.

PURPOSE: To develop a model integrating Single Photon Emission Computed Tomography (SPECT) radiomic features and clinical variables for predicting the therapeutic efficacy of 131I treatment in patients with DTC.

MATERIALS AND METHOD: The SPECT images and clinical data from 311 patients were included in this study. A total of 1,688 radiomic features were extracted from SPECT images. We investigated the diagnostic performance of 30 cross-combined models. Radiomic features were then combined with clinical parameters using early-fusion and late-fusion strategies. To further enhance predictive performance, an input sequence reflecting the selected features and the prediction task was designed and fed into a selected Large Language Model (LLM), which was fine-tuned using low-rank adaptation to optimize fused feature representations. Model performance was evaluated using accuracy (Acc), sensitivity (Sens), positive predictive value (PPV), F1-score, and the area under the curve (AUC).

RESULTS

The Relief-F + random forest radiomic model achieved the highest AUC (0.69). Multimodal models combining radiomics with clinical features outperformed clinical-only models, demonstrating the added predictive value of radiomics. The final LLM-enhanced multimodal model achieved the best performance, with an AUC of 0.85 and notable improvements in sensitivity and F1-score.

CONCLUSIONS

The proposed multimodal framework provides accurate early prediction of RAI therapeutic response in DTC and holds promise for improving individualized treatment planning and follow-up strategies.