CRViT-YOLO: A method for multi-morphological blood cell detection using convolution-restructured vision transformer.

Area of Science:

• Medical Diagnostics
• Computer Vision
• Artificial Intelligence

Background:

• Manual complete blood cell counting is laborious and error-prone.
• Deep learning offers a promising avenue for automated blood cell analysis.
• Existing methods face challenges with cell morphology, staining, and image quality variations.

Purpose of the Study:

• To develop an advanced deep learning framework for accurate and efficient blood cell detection.
• To improve feature extraction and localization accuracy in blood cell images.
• To validate the proposed model's performance on diverse blood cell datasets.

Main Methods:

• Developed CRViT-YOLO, a novel detection framework based on the YOLOv9 architecture.
• Incorporated a Convolutional-Reconstructed Vision Transformer (CRViT) module for enhanced feature extraction.
• Integrated a Feature Enhancement Module (FEM) and EIoU loss function for refined feature representation and localization accuracy.

Main Results:

• CRViT-YOLO achieved high mean average precision (mAP@50) scores on four public datasets: BCCD (93.9%), BCDD (99.4%), LISC (98.8%), and BBBC041 (76.0%).
• The model demonstrated robust performance in detecting polymorphic, healthy, and pathological cells.
• Effectively handled densely packed or overlapping cells across various scales and types.

Conclusions:

• CRViT-YOLO presents a highly effective and generalizable solution for automated multi-class blood cell detection.
• The framework significantly enhances accuracy and efficiency compared to conventional methods.
• The proposed approach shows strong potential for clinical diagnostic applications.