SCKansformer: Fine-Grained Classification of Bone Marrow Cells via Kansformer Backbone and Hierarchical Attention Mechanisms

TL;DR

Bone marrow cell classification is essential for diagnosing hematologic diseases but is hampered by high-dimensional micrograph data, long-tail class distributions, and subtle inter-class differences. The authors introduce SCKans transformer, which fuses a Kansformer Encoder (replacing MLP with Kolmogorov-Arnold networks), an SCConv Encoder (spatial and channel redundancy reduction), and a Global-Local Attention Encoder (global self-attention plus local feature extraction) to achieve robust fine-grained classification. The model is validated on a private BMCD-FGCD dataset (>10k samples, ~40 classes) and public BM datasets (PBC, ALL-IDB), outperforming ViT, EfficientNetV2, and specialized WBC models across accuracy, precision, recall, F1, and MCC; ablation studies confirm the necessity of each component. Additionally, the BMCD-FGCD dataset is released to the research community, underscoring the method’s practical impact for hematology and automated diagnostics and setting the stage for broader clinical deployment in bone marrow cytomorphology.

Abstract

The incidence and mortality rates of malignant tumors, such as acute leukemia, have risen significantly. Clinically, hospitals rely on cytological examination of peripheral blood and bone marrow smears to diagnose malignant tumors, with accurate blood cell counting being crucial. Existing automated methods face challenges such as low feature expression capability, poor interpretability, and redundant feature extraction when processing high-dimensional microimage data. We propose a novel fine-grained classification model, SCKansformer, for bone marrow blood cells, which addresses these challenges and enhances classification accuracy and efficiency. The model integrates the Kansformer Encoder, SCConv Encoder, and Global-Local Attention Encoder. The Kansformer Encoder replaces the traditional MLP layer with the KAN, improving nonlinear feature representation and interpretability. The SCConv Encoder, with its Spatial and Channel Reconstruction Units, enhances feature representation and reduces redundancy. The Global-Local Attention Encoder combines Multi-head Self-Attention with a Local Part module to capture both global and local features. We validated our model using the Bone Marrow Blood Cell Fine-Grained Classification Dataset (BMCD-FGCD), comprising over 10,000 samples and nearly 40 classifications, developed with a partner hospital. Comparative experiments on our private dataset, as well as the publicly available PBC and ALL-IDB datasets, demonstrate that SCKansformer outperforms both typical and advanced microcell classification methods across all datasets. Our source code and private BMCD-FGCD dataset are available at https://github.com/JustlfC03/SCKansformer.

Figures (6)

• Figure 1: The overall architecture of our proposed SCKansformer model. The SCKansformer model primarily comprises three parts: Kansformer Encoder, SCConv Encoder and Global-Local Attention Encoder.
• Figure 2: The framework of Global-Local Attention Encoder. The Global-Local Attention Encoder combines the MSA module and the Local Part module to effectively capture the global and local features of microscopic images, which enhances the model's ability to recognize long-distance dependencies and fine-grained features.
• Figure 3: The framework of SCConv Encoder. The SCConv Encoder optimizes the feature representation by means of Spatial Reconstruction Units and Channel Reconstruction Units, which reduces feature redundant information of microscopic images.
• Figure 4: The feature dimensionality reduction visualizations of SCConv Encoder and other traditional methods. (a) Heatmap of PCA. (b) Heatmap of Autoencoder. (c) Heatmap of SCConv Encoder.
• Figure 5: (a) Statistical presentation on the number of different types of bone marrow blood cells in our BMCD-FGCD dataset. Our dataset is divided according to categories, the training set contains 73877 samples and the testing set contains 18458 samples, and the ratio of the training set to the testing set is 8:2. (b) Typical cell body images of different types of bone marrow blood cells in our BMCD-FGCD dataset. Intuitive visualization of information such as morphological features, color and structure of various types of bone marrow blood cells.