Enhancing Morphological Discrimination in Red Blood Cell Images via Multi-Level Attention
Year: 2025
Type: conference-paper
Abstract: Reliable red blood cell classification is essential for facilitating early detection and treatment of hematological diseases. Nonetheless, the considerable morphological diversity among cell types presents a major challenge for traditional convolutional neural networks. In response to this issue, we introduce MAFPN (Multi-scale Attention Feature Pyramid Network), an attention-integrated architecture built on a ResNet-based FPN framework. By embedding attention modules at multiple hierarchical levels, MAFPN effectively enhances feature representations by emphasizing discriminative patterns and filtering out irrelevant signals. Evaluations on a four-category red blood cell dataset indicate that our method consistently surpasses the baseline FPN across various standard performance metrics. These findings suggest that MAFPN constitutes a promising solution for fine-grained medical image classification, leveraging attention-guided multi-scale representation learning.
Institution: University of Toyama
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W7128364711
