A Generative Multi-modal Deep Learning Framework for Non-invasive Prediction of Ki-67 Expression in Papillary Thyroid...

Accurate assessment of Ki-67 expression plays a critical role in the risk stratification and prognosis of papillary thyroid microcarcinoma (PTMC). However, its clinical evaluation still relies on invasive biopsy. To address this limitation, we propose a novel generative multi-modal and multi-task learning framework for non-invasive prediction of Ki-67 expression using only preoperative B mode ultrasound. Specifically, a BUS2USE network based on Pix2Pix is trained on paired B-mode and elastography images to synthesize virtual ultrasound elastography, thereby providing complementary tissue stiffness information without requiring specialized equipment. To mitigate the potential noise introduced by synthetic images for diagnosis, we design an reliability-aware fusion mechanism that dynamically balances the contributions of real and generated features, which is further enhanced by a multi-scale attention refinement module (MARM) for robust feature representation. Additionally, we introduce VEGF prediction as an auxiliary task within a multi-task learning framework, using biologically related angiogenesis information as complementary supervision to regularize Ki 67-oriented feature learning. Extensive experiments demonstrate that the proposed method outperforms existing CNN and transformer-based models, achieving an accuracy of 0.852 and an AUC of 0.889. Additional auxiliary label control and degradation robustness analyses further support the reliability of the proposed framework. The results validate the effectiveness of integrating synthetic imaging, attention-based fusion, and multi-task supervision for non-invasive, real-time Ki-67 prediction. This approach is particularly valuable for small nodules where invasive testing is not routinely recommended.