A Multi-Stage Deep Learning Framework For Lymphoma Classification And Morphology-Aware Follicular Lymphoma Grading

Abstract

Accurate diagnosis of lymphomas is difficult as some of the lymphomas are morphologically similar and the grading of follicular lymphomas (FL) is variable. This study proposes an automatic hierarchical framework to identify lymphoma, classify its type and grade it according to the World Health Organization (WHO) classification of lymphomas from histopathological images. However, a named benchmark dataset is not indicated in the provided implementation, the experimental images include benign tissue, Chronic Lymphocytic Leukemia (CLL), Follicular Lymphoma (FL), Mantle Cell Lymphoma (MCL) and various grades of FL. Both images are resized to 224 × 224 and processed by two different classifiers: a binary deep-learning classifier trained to differentiate between benign and lymphoma, and a multiclass deep-learning classifier for the classification of CLL, FL and MCL. In FL cases, the FollicularLymphomaGrader algorithm is based on the combination of nucleus segmentation, adaptive morphological analysis, estimation of circularity, detection of nucleoli, characterization of centroblasts–centrocytes and WHO grading criteria. The classification demonstrated here had a 96.1% confidence level in detecting lymphoma, and 100.0% confidence in recognizing FL. Seven centroblasts and 120 centrocytes were detected at the cell level and the ratio was 0.06 CB:CC. In particular, the novel integration of hierarchical deep learning with cell-level, interpretable WHO grading in a single diagnostic pipeline is the key novelty. The suggested framework can offer automatic disease classification, identification of the lymphoma subtypes, quantitative assessment of the cells, and a new clinically interpretable grading, thus offering a good computer-aided solution for histopathological diagnosis of lymphoma.