MACHINE LEARNING IN THE DIAGNOSIS OF FOLLICULAR LYMPHOMA

Abstract

Objective To implement a machine-learning (ML) model to assist pathologists in the differentiation of follicular hyperplasia (FH) and follicular lymphoma (FL). Study Design Whole slide images from 10 patients with FH and 10 patients with FL were manually annotated and fragmented into 21,585 patches (FH=13,399 and FL=8,186) of 299 × 299 pixels. The convolutional neural network (CNN) VGGNet was re-trained using Python 3.6 and other open-source libraries for machine learning and image processing (TensorFlow, Keras, Scikit-Learn, and OpenCV). The training and validation were carried out for 10 epochs until accuracy stabilized and validation loss reduced its variation. Results The total processing time for the CNN training was 753s. Different metrics could be obtained through the confusion matrix, emphasizing a high training accuracy of 98% and F1-score of 82%. Sensitivity and specificity were 74.8% and 91.8%, respectively. The receiver operating characteristic curve of 94% showed the fine class separation ability of the CNN. Conclusion The ML model used in this study is feasible to differentiate FH and FL. Additional CNN training and validation in bigger/multicenter datasets may generate AI-assisted tools to aid FL diagnosis. To implement a machine-learning (ML) model to assist pathologists in the differentiation of follicular hyperplasia (FH) and follicular lymphoma (FL). Whole slide images from 10 patients with FH and 10 patients with FL were manually annotated and fragmented into 21,585 patches (FH=13,399 and FL=8,186) of 299 × 299 pixels. The convolutional neural network (CNN) VGGNet was re-trained using Python 3.6 and other open-source libraries for machine learning and image processing (TensorFlow, Keras, Scikit-Learn, and OpenCV). The training and validation were carried out for 10 epochs until accuracy stabilized and validation loss reduced its variation. The total processing time for the CNN training was 753s. Different metrics could be obtained through the confusion matrix, emphasizing a high training accuracy of 98% and F1-score of 82%. Sensitivity and specificity were 74.8% and 91.8%, respectively. The receiver operating characteristic curve of 94% showed the fine class separation ability of the CNN. The ML model used in this study is feasible to differentiate FH and FL. Additional CNN training and validation in bigger/multicenter datasets may generate AI-assisted tools to aid FL diagnosis.