Small Object Detection via Pixel Level Balancing With Applications to Blood Cell Detection.

Area of Science:

• Medical Imaging Analysis
• Computer Vision in Healthcare
• Diagnostic Technologies

Background:

• Object detection is crucial for medical diagnosis, aiding in tasks like blood cell analysis.
• Detecting small, influential objects in medical images presents significant challenges.
• Existing methods focus on feature extraction and data augmentation, but overlook inter-object size interference during training.

Purpose of the Study:

• To address the challenge of small object detection in medical imaging.
• To propose a novel method, pixel level balancing (PLB), to mitigate interference between objects of different sizes during model training.
• To enhance the accuracy of small object detection for improved medical diagnostic assistance.

Main Methods:

• Introduced pixel level balancing (PLB) as a novel approach for object detection.
• Characterized object size using the number of pixels within the detection box.
• Dynamically adjusted the training loss of objects based on their size using a weighting mechanism.

Main Results:

• Demonstrated that objects of different sizes interfere with each other during the training of detection models.
• Showcased that the proposed pixel level balancing (PLB) method effectively improves the accuracy of small object detection.
• Validated the method's performance, particularly in blood cell detection scenarios.

Conclusions:

• Object size disparity is a significant factor affecting detection model performance.
• Pixel level balancing (PLB) offers a viable solution to enhance small object detection accuracy in medical applications.
• The PLB method shows promise for improving diagnostic capabilities through more precise medical image analysis.