A hierarchical deep learning model based on whole slide imaging of cerebrospinal fluid cells for rapid diagnosis of meningeal carcinomatosis
10.3760/cma.j.cn114452-20250225-00109
- VernacularTitle:开发一种基于全玻片数字图像的脑脊液细胞分层深度学习模型用于诊断脑膜癌
- Author:
Kun CHEN
1
;
Xiangyu LI
;
Qianqian XU
;
Zhiyu XU
;
Di WANG
;
Huanhuan QIN
;
Guangjie JIANG
;
Haoqin JIANG
;
Qiong ZHAN
;
Mengxi GE
;
Xin LI
;
Chun XU
;
Ming GUAN
Author Information
1. 复旦大学附属华山医院检验医学科，上海 200040
- Publication Type:Journal Article
- Keywords:
Whole slide image;
Cerebrospinal fluid;
Deep learning;
Convolutional neural networks;
Meningeal carcinomatosis;
Tumor cell
- From:
Chinese Journal of Laboratory Medicine
2025;48(12):1558-1564
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a convolutional neural network model of whole slide imaging of cerebrospinal fluid cells for rapid and accurate identification and classification of tumor cells in cerebrospinal fluid.Methods:A total of 8 692 cerebrospinal fluid cytology smears from Huashan Hospital Affiliated to Fudan University from January 2nd, 2019, to December 27th, 2024. As randomly assigned, the training set included 4 941 benign and 1 745 malignant samples, while the validation set comprised of 1 368 benign and 638 malignant samples. Whole-slide digital images were acquired using a cytopathology scanner, cells (clusters) were annotated for classification, and a deep learning model was constructed via tiled image patches for cell detection and classification. Model performance was evaluated using accuracy, sensitivity, specificity, and other indicators. The classification efficiency of manual microscopy was compared.Results:The model achieved a mean precision of 96.75% for cerebrospinal fluid cell classification. For malignant tumor cells, the classification accuracy was 96.61% (mAP=98.36%, AUC=0.97). Subtype classification accuracies for epithelial/epithelioid tumors and small round cell tumors were 97.13% (AUC=0.98) and 95.58% (AUC=0.93), respectively. Compared with manual microscopy, which took (9.70±0.82) minutes for classifying 200 cells, (18.27±1.21) minutes for 500 cells, and often exceeded 60 minutes or infeasible for full slides, the AI model took (3.46±0.49) seconds for 200 cells, (6.76±0.82) seconds for 500 cells, and a median of 48.57 seconds for full slides ( P<0.001), representing an efficiency improvement of approximately 161-170 times, significantly enhancing diagnostic efficiency. Conclusion:This fully automated hierarchical deep learning model enables efficient and accurate tumor cell identification and classification in CSF, providing an effective auxiliary tool for the rapid diagnosis of meningeal carcinomatosis.
