A DenseNet-121-Based Multi-Scale Attention Fusion Framework for Peripheral Blood Cell Classification

Abstract

Automated peripheral blood cell classification is significant in computer-aided hematological diagnosis. However, the extreme class imbalance and morphological resemblance of leukocyte subtypes pose obstacles to accurate classification. In this paper, a custom DenseNet-121 architecture with a lightweight multi-scale attention-based feature fusion module is proposed for peripheral blood cell classification. Instead of using only the last feature representation as in traditional DenseNet-121 models, our proposed method extracts intermediate feature maps from several backbone stages and adaptively combines them with scale-aware attention weighting. Then the fused representation is refined for classification by channel-wise attention and spatial attention mechanisms. Experiments are performed on the publicly available KU-Optofil peripheral blood cell dataset which contains 31,489 microscopic blood cell images of 13 classes. A systematic grid search involving 48 hyperparameter configurations was performed to determine the optimal training setup. To address class imbalance, minority classes in the training set were augmented up to the size of the largest training class, while validation and test sets were kept unchanged. The impact of red-green-blue, hue-saturation-value, perceptually uniform color, and grayscale representations was also explored. The best result was obtained with hue-saturation-value representation, reaching 96.31% test accuracy and 84.78% macro-averaged harmonic mean of precision and recall. These results indicate that combining multi-scale attention fusion, minority-class augmentation, and color space selection can improve macro-level recognition performance without changing the underlying network structure.