A hybrid deep learning framework for platelet detection and counting in blood smear images toward dengue-relevant applications

Accurate platelet quantification is essential for diagnosing dengue, thrombocytopenia, leukemia, and other hematological disorders; however, conventional manual microscopy and automated hematology analyzers often suffer from reduced reliability due to small platelet size, overlapping cells, staining variability and inconsistent illumination. To address these challenges, this study presents a deep learning–based framework for platelet detection, segmentation and counting using publicly available blood smear datasets, namely the BCCD Blood Cell Detection dataset and the Peripheral Blood Cell Image (PBCI) dataset. The proposed approach combines a lightweight super-resolution enhancement module with a Mask2Former transformer-based instance segmentation architecture, specifically optimized for platelet-scale object recognition. Super-resolution enhancement improves platelet visibility and preserves fine structural details, while multi-scale masked attention enables accurate instance segmentation in dense and morphologically complex smear regions. In addition, a density-map regression head is incorporated to improve counting reliability in highly congested areas. The framework was evaluated using strictly separated training, validation, and test subsets under a fixed experimental configuration, with the test data isolated from model selection and hyperparameter optimization. Quantitative evaluation demonstrates that the proposed framework consistently outperforms conventional CNN-based and transformer-only baselines. Compared with representative CNN detectors, the method achieves an improvement of 8–16% in AP_small and a 38–52% reduction in mean absolute counting error across the datasets. These results confirm that joint resolution enhancement and masked-attention–based segmentation improves platelet detection and counting accuracy under challenging smear conditions.