Exploring Novel Deep Learning Frameworks for Blood Cell Classification: From Algorithmic Design to Data Enhancement and Efficient Adaptation

Deep learning has been widely adopted in medical image analysis due to its ability to learn informative representations. However, its application to blood cell classification remains challenging. In particular, the limited availability and low quality of annotated blood cell datasets can lead to overfitting when training parameter-heavy models, thus hindering generalization and downstream diagnostic utility. Blood cell analysis presents unique challenges, including high inter-class similarity, subtle morphological differences, and a high dependency on expert annotations. These factors make the task particularly sensitive to overfitting and call for domain-specific adaptations. This thesis aims to address the issue of data scarcity and model overfitting in blood cell classification through a comprehensive strategy spanning algorithm design, data-level enhancement, and practical deployment. At the algorithmic perspective, a parameter-efficient model named ReRNet is proposed, in which the reliance on large pre-trained backbones is reduced by incorporating lightweight randomized neural layers. At the data perspective, generative adversarial network (DLBCNet) are proposed to synthesize realistic blood cell images, mitigating the lack of diversity and annotation in existing datasets. At the application level, an adapter-based method (LKA) with large kernel attention is introduced to effectively capture global contextual features while keeping computational cost manageable. The proposed methods are evaluated not only on blood cell datasets but also validated across other medical imaging modalities to demonstrate broad adoption. Overall, this work presents a targeted approach to improving the reliability and applicability of deep models in low resource medical scenarios, with a specific focus on blood cell classification.