Improved YOLOv7 Algorithm for Detecting Bone Marrow Cells.

Area of Science:

• Hematology
• Computational Pathology
• Medical Imaging Analysis

Background:

• Accurate bone marrow (BM) cell detection and classification are vital for hematology diagnosis.
• Current methods face challenges due to limited data, subtle class differences, and small target sizes, leading to low accuracy and significant manual workload for pathologists.
• There is a need for automated, high-accuracy solutions to improve diagnostic efficiency and reliability.

Purpose of the Study:

• To develop an improved BM-cell detection algorithm, YOLOv7-CTA, to address the limitations of existing methods.
• To enhance the model's sensitivity to fine-grained features and improve attention to small target cells.
• To optimize anchor box generation and handle class imbalance for better detection performance.

Main Methods:

• Proposed YOLOv7-CTA algorithm incorporating a novel CoTLAN module for long-term feature modeling in the backbone network.
• Integrated coordinate attention (CoordAtt) module to enhance focus on small target features.
• Utilized K-means++ clustering for generating optimized anchor boxes and Focal loss to address positive-negative sample imbalance.

Main Results:

• The YOLOv7-CTA model achieved a best mean average precision (mAP) of 88.6%.
• Demonstrated significant improvements over existing models: 12.9% over Faster R-CNN, 8.3% over YOLOv5l, and 6.7% over YOLOv7.
• Effectively addressed challenges related to subtle feature differences, small target sizes, and imbalanced datasets.

Conclusions:

• The YOLOv7-CTA model shows superior effectiveness and performance in bone marrow cell detection tasks.
• The developed algorithm offers a promising automated solution for improving accuracy and efficiency in hematological diagnostics.
• The integration of novel modules and optimization techniques significantly advances the state-of-the-art in medical image analysis for cell classification.