深度学习辅助痰涂片抗酸杆菌分割模型构建与多中心验证

Deep Learning-Assisted Construction and Multicenter Validation of an AFB Segmentation Model on Sputum Smears

摘要: 目的：构建并验证一种面向痰涂片抗酸杆菌(Acid-Fast Bacilli, AFB)像素级识别的深度学习分割模型，评价其在多中心场景下的泛化能力、稳定性与临床可应用性。方法：基于4家医疗中心319例患者数据，纳入5647张含目标像素级标注图像块，采用分阶段消融策略开展模型构建与优化。在统一训练框架下，依次比较多中心训练策略、负样本混入比例、编码器容量、注意力机制和输入分辨率；通过同中心测试、多中心验证、5折交叉验证及留一中心交叉验证(LOCO)进行综合评估。结果：最优配置为U-Net + ResNet50 + scSE + DiceBCE + Strong增强 + 512 × 512，测试集Dice为0.8760，IoU为0.7794，多中心Dice为0.8579。5折交叉验证Dice为0.8741 ± 0.0019，LOCO Dice为0.8591 ± 0.0124，泛化差距为1.50%。数据策略显示，简单多中心混合训练未形成稳定增益；负样本按1:1~1:3比例直接混入会降低泛化性能。单张图像推理约0.18 s。结论：在多中心痰涂片AFB分割任务中，以系统消融驱动的模型构建路径可同时获得较高精度、较低泛化差距和可接受推理效率，可为结核病实验室镜检流程的标准化与智能化提供技术支持。

Abstract: Objective: To develop and validate a deep learning–based segmentation model for pixel-level recognition of Acid-Fast Bacilli (AFB) in sputum smear images, and to evaluate its generalization ability, robustness, and clinical applicability across multi-center settings. Methods: Based on data from 319 patients across four medical centers, a total of 5,647 annotated image patches containing target bacilli were collected. A staged ablation strategy was adopted for model construction and optimization. Within a unified training framework, we systematically compared multi-center training strategies, the proportion of negative sample mixing, encoder capacity, attention mechanisms, and input resolution. Model performance was comprehensively evaluated through intra-center testing, multi-center validation, five-fold cross-validation, and Leave-One-Center-Out (LOCO) validation. Results: The optimal configuration (U-Net + ResNet50 + scSE + DiceBCE + Strong Augmentation, 512 × 512) achieved a Dice coefficient of 0.8760 and an IoU of 0.7794 on the test set. Multi-center Dice was 0.8579. Five-fold cross-validation yielded a Dice of 0.8741 ± 0.0019, while LOCO validation achieved 0.8591 ± 0.0124, with a performance drop of less than 1.50%. Data strategy analysis showed that simple multi-center data mixing did not yield stable improvements, and incorporating negative samples at a ratio of 1:1 - 1:3 reduced model performance. The inference time per image was approximately 0.18 s. Conclusion: In multi-center AFB sputum smear segmentation tasks, a systematically ablation-driven model development strategy can achieve high accuracy, minimal generalization loss, and acceptable inference efficiency, thereby providing technical support for the standardization and intelligent transformation of laboratory diagnostic workflows for tuberculosis.