Classification of Digital Pathological Images of Non-Hodgkin's Lymphoma Subtypes Based on the Fusion of Transfer Learning and Principal Component Analysis.

PURPOSE Non-Hodgkin's lymphoma (NHL) is a serious malignant disease. Delayed diagnosis will cause anemia, increased intracranial pressure, organ failure, and even lead to death. The current main trend in this area is to use deep learning (DL) for disease diagnosis. Extracting classification information from the digital pathology images by DL may realize the automated qualitative and quantitative analysis of NHL. Previously, DL has been used to classify NHL digital pathology images with some success. However, shortcomings still exist in the data preprocessing methods and feature extraction. Therefore, this paper presents a method for the classification of NHL subtypes based on the fusion of transfer learning (TL) and principal component analysis (PCA). METHODS First, the NHL digital pathology images were preprocessed by image division and segmentation and then input into the transfer models for fine-tuning and feature extraction. Second, PCA was used to map the extracted features. Finally, a neural network was used as a classifier to classify the mapped features. During the fine-tuning of the transfer models, two methods, freezing all feature extraction layers and fine-tuning all layers, were employed to select the optimal model with the best classification result among all the pre-selected transfer models. On this basis, the use of freezing the layers' location was discussed and analyzed. RESULTS The results show that the proposed method achieved average 5-fold cross-validation accuracies of 100%, 99.73%, and 99.20% for chronic lymphocytic leukemia (CLL), follicular lymphoma (FL), and mantle cell lymphoma tumor (MCL), and each category has standard deviations 0.00, 0.53, and 0.65, respectively, in the NHL reference dataset. The overall classification accuracy for 5-fold cross-validation is 98.93%, which is an increase of 1.26% compared to the latest reported methods, having a lower standard deviation (1.00). CONCLUSION The method proposed in this paper achieves a high classification accuracy and strong model generalization for the classification of NHL, which makes it possible to conduct intelligent classification of NHL in clinical practice. Our proposed method has definite clinical value and research significance.