Explainable Classifier for Malignant Lymphoma Subtyping via Cell Graph and Image Fusion

TL;DR

The paper tackles malignant lymphoma subtyping from whole slide images by introducing an explainable multimodal MIL framework that jointly leverages ROI-based image regions and cell-distribution patterns captured via labeled cell graphs. By pretraining AdditiveMIL for each modality and then fusing them with a Weak-Expert-based Gating MoE, the approach provides per-patch ROI attention and cell-type frequency/spatial distribution explanations while achieving state-of-the-art accuracy on a 1,233-WSI dataset across DLBCL, FL, and Reactive subtypes. Key contributions include the integration of cell graphs and image features for MIL, explainable class-wise attention, and first application of cell-graph analysis to malignant lymphoma, validated against pathologist ROIs. The results demonstrate both superior performance and explanations that align with clinical reasoning, suggesting practical potential for clinical adoption.

Abstract

Malignant lymphoma subtype classification directly impacts treatment strategies and patient outcomes, necessitating classification models that achieve both high accuracy and sufficient explainability. This study proposes a novel explainable Multi-Instance Learning (MIL) framework that identifies subtype-specific Regions of Interest (ROIs) from Whole Slide Images (WSIs) while integrating cell distribution characteristics and image information. Our framework simultaneously addresses three objectives: (1) indicating appropriate ROIs for each subtype, (2) explaining the frequency and spatial distribution of characteristic cell types, and (3) achieving high-accuracy subtyping by leveraging both image and cell-distribution modalities. The proposed method fuses cell graph and image features extracted from each patch in the WSI using a Mixture-of-Experts (MoE) approach and classifies subtypes within an MIL framework. Experiments on a dataset of 1,233 WSIs demonstrate that our approach achieves state-of-the-art accuracy among ten comparative methods and provides region-level and cell-level explanations that align with a pathologist's perspectives.

Figures (3)

• Figure 1: Left to right: DLBCL, FL, and Reactive case, and cells characteristic of each subtype (DLBCL corresponds to LBC, FL to CC, and Reactive to RM). Red lines indicate ROIs, and green and red lines in Reactive and FL, respectively, indicate follicles. Heat maps indicate the spatial distribution of the number of cells characteristic of each subtype in a 512-pixel square at 40x magnification.
• Figure 2: (A) Annotations on a t-SNE map to label cells with LBC, CC, RM, or others. (B) Example of a cell graph to be constructed. The nodes' color indicates the cell type. (C) An overview of our method, WEG-MoE, for classifying the $n$-th WSI.
• Figure 3: (A) Class-wise attention, where higher attention is red, and lower is blue. (B) Frequency by cell types. (C) Frequency of adjacency between cell types. In (B) and (C), the black dotted lines present the results computed with the ROI supervised by a pathologist, which can be interpreted as the ground truth, and the orange dotted lines present the median of the distribution. In (C), for example, "LBC$\to$CC" indicates the number of CCs that are connected by edges to LBC. Each data in (B) and (C) is from the top 25% of instance-level class-wise attention scores for the correct class.