ORF40-03 - Thyroid Tumor Score-an AI-Driven Multimodal Model Combining Ultrasound and Multiplex Molecular Cytology for Risk Stratification of Thyroid Nodules

Thyroid Tumor Score—An AI-Driven Multimodal Model Combining Ultrasound and Multiplex Molecular Cytology for Risk Stratification of Thyroid Nodules

Background: Indeterminate thyroid nodules, based on the current Bethesda system, often require diagnostic surgery, risking inappropriate treatment and long-term morbidity in a significant number of patients. While integrating Bethesda and TIRADS scores improves diagnostic yield, subjectivity remains a challenge. The current study presents a multimodal, Artificial Intelligence (AI)-driven framework that combine ultrasound (US) and marker-based cytology data to enhance diagnostic precision.

Methods: USG-guided FNAC was performed on 697 nodules. Bethesda II and Bethesda V/VI nodules were considered benign and malignant, respectively. Bethesda III/IV nodules that underwent surgery with definitive histopathology were included for analysis. Multiplex molecular cytochemistry was performed on FNAC samples using Galectin-3 and wheat germ agglutinin (WGA). U-Net–based models were trained for segmentation of nodules from USG images and cytology thyroid clusters/cells. A multimodal Thyroid Tumor Score (TTS) was developed using a Weakly-Supervised Attention model by pairing cytology “bags” (segmented clusters/cells) with the corresponding segmented USG nodule images.

Results: Of 637 participants and 697 eligible nodules, 448 were included in the final analysis. Most were Bethesda II (n=291; surgery= 19; 15 Benign) and Bethesda V/VI (n=93; surgery 55; malignant =53). Among Bethesda III/IV nodules with definitive diagnosis (n=63), 37 were benign and 26 malignant. For USG nodule segmentation, U-Net achieved an IoU of 0.90 (training images=2725; test images n=231). USG classification using ResNet34 (358 nodules; training images benign=14,136 and malignant=5,916) achieved a specificity of 85%. Cytology segmentation achieved a cluster-level IoU of 0.74 (training images n=540; test images n=131), with cluster classification accuracy of 78% (Benign =1707, malignant =1905 images) using ResNet34.
In the multimodal weakly supervised attention model (training nodules, n=343; benign=258, malignant=85), overall accuracy was 89% (accurately classified= 38/43) for benign vs malignant classification. For indeterminate nodules, performance was 100% sensitivity (9/9) and 82% specificity (9/11). Attention maps showed high-attention regions over thyroid clusters and corresponding USG regions in malignant cases.

Conclusion: This multimodal AI framework combining USG and marker-based cytology demonstrates strong performance for thyroid nodule stratification and shows particular promise for improving triage of indeterminate nodules.