目的 评价一种基于深度学习的两阶段近红外自体荧光（NIRAF）甲状旁腺识别模型在提高手术中甲状旁腺识别准确性方面的效果，并验证其在降低假阳性率和提供解剖学定位参考中的应用价值。方法 前瞻性纳入福建医科大学附属协和医院2023年1月至2023年12月期间接受甲状腺切除术的101例病人。术后在标准手术间灯光下、固定距离15 cm处使用785 nm激光近红外相机采集新鲜标本NIRAF影像，共获得30 122帧图像，构建Niraf24数据集。模型由YOLOv8x检测网络和Segment Anything Model v2（SAM2）分割网络组成，前者用于荧光信号定位，后者用于甲状腺轮廓分割及背景恢复。通过灵敏度（sensitivity）、精确率（precision）、F1值、假阳性率及交并比（IoU）等指标评价模型性能，并与传统荧光直读法比较。结果 两阶段模型显著降低了假阳性比例，帧水平由37.7%降至13.2%，标本水平由38.8%降至6.7%（均P<0.001），精确率由62.3%升至86.8%，F1值由0.734升至0.846，灵敏度保持在82.8%。SAM2生成的甲状腺轮廓平均交并比为（0.985±0.012），显示其解剖定位高度一致。模型在高噪声环境（棕色脂肪、热凝痂、弥漫性高荧光背景、手术染料污染等）下依然保持稳定性能，显著抑制困难假阳性信号。在不降低灵敏度的前提下，两阶段级联相较传统判读显著降低假阳性并提升整体效能，验证了“先检后分”的任务适配性与术中应用可行性。结论 基于YOLOv8-SAM2的两阶段深度学习级联模型在不降低灵敏度的前提下显著提高了NIRAF的特异性，能够实时生成具有解剖学参照意义的可视化信息，为甲状旁腺保护提供技术支持。该方法建立了可重复的性能基准，为多中心临床验证和术中智能决策提供了新的思路，有助于减少甲状旁腺误切与术后低钙血症的发生。

Objective To assess the effectiveness of a two-stage deep-learning near-infrared autofluorescence (NIRAF) model for improving intraoperative parathyroid identification accuracy and to validate its utility in reducing false-positive rates and providing anatomical localization reference. Methods A total of 101 patients undergoing thyroidectomy at Fujian Medical University Union Hospital from January 2023 to December 2023 were prospectively enrolled. Under standard operating-room lighting at a fixed distance of 15 cm, a 785-nm laser NIR camera was used to acquire NIRAF images from fresh specimens’ post-resection, yielding 30,122 frames to construct the “Niraf24” dataset. The model consisted of a YOLOv8x detection network for fluorescence-signal localization and a Segment Anything Model v2 (SAM2) segmentation network for thyroid contour segmentation and background restoration. Performance was evaluated using sensitivity, precision, F1-score, false-positive rate, and intersection over union (IoU), and compared with conventional direct fluorescence reading. Results The two-stage model markedly reduced false positives: frame-level from 37.7% to 13.2% and specimen-level from 38.8% to 6.7% (both P<0.001). Precision increased from 62.3% to 86.8%, F1-score improved from 0.734 to 0.846, while sensitivity was maintained at 82.8%. The SAM2-derived thyroid contours achieved a mean IoU of (0.985±0.012), indicating highly consistent anatomical localization. The model remained robust in high-noise settings (brown adipose tissue, thermal coagulation eschar, diffuse high-fluorescence background, surgical dye contamination), effectively suppressing difficult false-positive signals. Without compromising sensitivity, the two-stage cascade approach markedly reduced false positives and improved overall performance compared to conventional interpretation methods, validating the task suitability of the “detection-first, classification-second” strategy and its feasibility for intraoperative application. Conclusion The YOLOv8-SAM2-based two-stage deep-learning cascade significantly enhances the specificity of NIRAF without compromising sensitivity, generating real-time, anatomically referenced visual information to support parathyroid preservation. This approach establishes a reproducible performance benchmark and offers a new paradigm for multicenter clinical validation and intraoperative intelligent decision-making, with the potential to reduce inadvertent parathyroidectomy and postoperative hypocalcemia.