BloodContourNet: A YOLO-Based Density-Aware Bézier Refinement Framework for Peripheral Blood Cell Detection

Area of Science:

• Medical Imaging
• Computational Biology
• Hematology

Background:

• Accurate peripheral blood cell detection is challenging in dense smear fields due to overlapping cells and staining variations.
• Existing box-based object detectors struggle with reliability in complex hematological images.
• Limitations include difficulty in distinguishing touching erythrocytes, small platelets, and staining inconsistencies.

Purpose of the Study:

• To develop BloodContourNet, a contour-informed and density-aware detection framework.
• To improve the localization accuracy of red blood cells, white blood cells, and platelets.
• To address the limitations of current detectors in dense peripheral blood smear fields.

Main Methods:

• BloodContourNet is based on the YOLOv26-L architecture.
• It integrates a Hematological Attention Module (HAM) for feature refinement.
• Includes a Density-Aware Reweighting Head (DARH) for loss modulation and a Bézier Contour Refinement Head (BCRH) for geometric cues.

Main Results:

• BloodContourNet achieved 99.1% mAP@IoU=50% and 87.8% mAP@50:95 on the TXL-PBC dataset.
• Demonstrated statistically significant improvements over YOLOv26-L and DINO-DETR in strict-IoU performance.
• Outperformed 25 YOLO variants and 10 Transformer-based detectors in benchmark experiments.

Conclusions:

• BloodContourNet enhances strict localization for challenging dense erythrocyte regions and small platelets.
• The framework offers a significant advancement over existing box-based detectors for peripheral blood cell analysis.
• Further external validation is required for broader clinical applicability.