A convolutional neural network based model for assisting pathological diagnoses on thyroid liquid-based cytology
10.3760/cma.j.cn112151-20200802-00613
- VernacularTitle:基于卷积神经网络的甲状腺液基细胞学病理辅助诊断模型的研究
- Author:
Meihua YE
1
;
Wanyuan CHEN
;
Bojun CAI
;
Chaohui JIN
;
Xianglei HE
Author Information
1. 杭州医学院附属人民医院 浙江省人民医院病理科，杭州 310014
- Keywords:
Artificial intelligence;
Thyroid diseases;
Cytological techniques
- From:
Chinese Journal of Pathology
2021;50(4):358-362
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To develop a convolutional neural network based model for assisting pathological diagnoses on thyroid liquid-based cytology specimens.Methods:Seven-hundred thyroid TCT slides were collected, scanned for whole slide imaging (WSI), and divided into training and test sets after labeling the correct diagnosis (benign versus malignant). The extracted regions of interest after noise filtering were cropped into pieces of 512 × 512 patch on 10 × and 40 × magnifications, respectively. A classification model was constructed using deeply learning algorithms, and applied to the training set, then automatically tuned in the test set. After data enhancement and parameters optimization, accuracy, sensitivity, specificity, positive predictive value and negative predictive value of the model were calculated. Results The training set with 560 WSI contained 4 926 cell clusters (11 164 patches), while the test set with 140 WSI contained 977 cell clusters (1 402 patches). YOLO network was selected to establish a detection model, and ResNet50 was used as a classification model. With 40 epochs training, results from 10× magnifications showed an accuracy of 90.01%, sensitivity of 89.31%, specificity of 92.51%, positive predictive value of 97.70% and negative predictive value of 70.82%. The area under curve was 0.97. The average diagnostic time was less than 1 second. Although the model for data of 40× magnifications was very sensitive (98.72%), but its specificity was poor, suggesting that the model was more reliable at 10× magnification. Conclusions:The performance of a deep-learning based model is equivalent to pathologists′ diagnostic performance, but its efficiency is far beyond. The model can greatly improve consistency and efficiency, and reduce the missed diagnosis rate. In the future, larger studies should have more morphology diversity, improve model′s accuracy and eventually develop a model for direct clinical use.
