DMPNet: Robust Real-Time Intraoperative Parathyroid Identification via Novel Dual-Modality Fusion Framework of White-Light and Spectral Imaging

Accurate intraoperative identification of parathyroid glands is crucial for preserving their function and min-imizing complications, such as hypocalcemia, during thyroid surgery. However, the small size, variable anatomical locations, and visual similarity of parathyroid glands to surrounding tissues make their identification challenging and heavily reliant on the surgeon’s expertise. To address these challenges, we propose DMPNet, a real-time dual-modality parathyroid detection framework that, for the first time, integrates white-light and spectral imaging for parathyroid detection. DMPNet is trained on a clinically relevant dataset comprising 3,309 white-light and 3,309 spectral images from 423 patients at Sun Yat-sen Memorial Hospital, reflecting diverse real-world surgical scenarios. The framework employs independent modality-specific feature extraction modules and a multi-scale fusion strategy to combine the complementary strengths of white-light and spectral imaging, significantly enhancing detection robustness and accuracy. The detection head processes these fused features, enabling precise localization and real-time identification of parathyroid glands. DMPNet achieves state-of-the-art (SOTA) performance, with a precision of 0.887, recall of 0.749, and an F1-score of 0.764 on the test set—outperforming existing parathyroid detection models. With an impressive inference speed of 350.6 frames per second (FPS), far exceeding the standard endoscopic video frame rate of 25 FPS, DMPNet provides real-time assistance to surgeons, reducing cognitive burden, enhancing surgical precision, and minimizing postoperative complications. This pioneering integration of white-light and spectral imaging represents a significant advancement in intraoperative parathyroid gland identification, offering transformative potential for improved patient outcomes.