Deep Learning for the Classification of Non-Hodgkin Lymphoma on Histopathological Images

Abstract

The diagnosis and the subtyping of non-Hodgkin lymphoma (NHL) are challenging and require expert knowledge, great experience, thorough morphological analysis, and often additional expensive immunohistological and molecular methods. As these requirements are not always available, supplemental methods supporting morphological-based decision making and potentially entity subtyping are required. Deep learning methods have been shown to classify histopathological images with high accuracy, but data on NHL subtyping are limited. After annotation of histopathological whole-slide images and image patch extraction, we trained and optimized an EfficientNet convolutional neuronal network algorithm on 84,139 image patches from 629 patients and evaluated its potential to classify tumor-free reference lymph nodes, nodal small lymphocytic lymphoma/chronic lymphocytic leukemia, and nodal diffuse large B-cell lymphoma. The optimized algorithm achieved an accuracy of 95.56% on an independent test set including 16,960 image patches from 125 patients after the application of quality controls. Automatic classification of NHL is possible with high accuracy using deep learning on histopathological images and routine diagnostic applications should be pursued.

Keywords: CLL/SLL; CNN; DLBCL; artificial intelligence; deep learning; histopathology.

Figure 1

Tumor annotation and generation of image patches. Representative tissue microarray core of a diffuse large B-cell lymphoma without (A) and with annotation (B)—yellow outline, as well as after image patche creation (C)—red squares. The image patches were subsequently saved as .png files. Tile numbers (Tile 1, Tile 2, etc.) are shown in a gray box within each red square in this example. Scale bars (black line): 100 µm.

Figure 2

Examples of image patches from annotated areas. Representative image patches from control lymph nodes (A), small lymphocytic lymphoma/chronic lymphocytic leukemia (B), and diffuse large B-cell lymphoma (C) are shown. Magnification: each image 100 × 100 µm (395 × 395 px).

Figure 3

Training and validation accuracy of the model with the highest validation accuracy as per EfficientNet.

Figure 4

Confusion matrix of the best-performing model in terms of the test data at the patch (left) and case level (right). The lower panels exhibit the balanced accuracy (BACC). CLL: chronic lymphocytic leukemia, DLBCL: diffuse large B-cell lymphoma, LN: lymph node.

Figure 5

SmoothGrad heatmaps of exemplary patches that were classified correctly. For each class, the upper plot shows the original image patch while the lower plot shows the patch overlaid with the SmoothGrad heatmap with respect to the class of the patch. High SmoothGrad activity scores can be seen in areas overlaid with single cells, as well as in lung LN with extracellular anthracosis. This confirms that the algorithm classified the image patches on the basis of cellular and extracellular morphological structures. DLBCL: diffuse large B-cell-lymphoma, LN: lymph node, SLL/CLL: small lymphocytic lympho-ma/chronic lymphatic leukemia. Magnification: each image 100 × 100 µm (395 × 395 px).