Neural Cellular Automata for Lightweight, Robust and Explainable Classification of White Blood Cell Images

TL;DR

Diagnosis of hematological malignancies relies on accurate white blood cell (WBC) identification from peripheral blood smears, but deep learning methods often struggle with domain shifts and lack of interpretability. We propose Neural Cellular Automata (NCA) as a lightweight, robust backbone for single WBC classification, extracting discriminative features through iterative local updates on a $64\times64$ grid and classifying with a small MLP. The system uses $n=128$ channels and $k=64$ update steps, achieving competitive accuracy with only about $86{,}000$ parameters across three datasets, and provides explainability via Layer-wise Relevance Propagation. These properties enable cross-domain robustness and potential deployment in clinical settings and low-resource environments.

Abstract

Diagnosis of hematological malignancies depends on accurate identification of white blood cells in peripheral blood smears. Deep learning techniques are emerging as a viable solution to scale and optimize this process by automatic cell classification. However, these techniques face several challenges such as limited generalizability, sensitivity to domain shifts, and lack of explainability. Here, we introduce a novel approach for white blood cell classification based on neural cellular automata (NCA). We test our approach on three datasets of white blood cell images and show that we achieve competitive performance compared to conventional methods. Our NCA-based method is significantly smaller in terms of parameters and exhibits robustness to domain shifts. Furthermore, the architecture is inherently explainable, providing insights into the decision process for each classification, which helps to understand and validate model predictions. Our results demonstrate that NCA can be used for image classification, and that they address key challenges of conventional methods, indicating a high potential for applicability in clinical practice.

Figures (4)

• Figure 1: Neural cellular automata (NCA) can be used for the accurate classification of single white blood cells in patient blood smears. A: Our approach consists of four steps: i) image padding to increase the number of channels, ii) $k$ NCA update steps to extract features from the image that manifest in the hidden channels, iii) pooling via channel-wise maximum, and iv) a fully connected network to classify the image. B: The NCA step updates each cell based on its immediate surroundings according to equations \ref{['eq:2']} - \ref{['eq:5']}. C: Training the model end-to-end allows the NCA to learn an update rule that extracts useful features.
• Figure 2: Class distribution for the three datasets.
• Figure 3: NCA classification accuracy saturates at 128 channels for two out of three white blood cell datasets.
• Figure 4: Example of ten features (columns) extracted by the NCA for an exemplary eosinophil, monocyte, neutrophil, and promyelocyte. The highest relevance scores have been attributed by layer-wise relevance propagation on the fully connected layers of the classifier network. The channel activations are located within the regions of the white blood cell indicating the model's ability to correctly identify and analyze pertinent aspects of the input data.