CSBC: Class and Scale Balanced Semi-Supervised Blood Cell Detection

Area of Science:

• Medical Imaging
• Computer Vision
• Hematology

Background:

• Accurate blood cell detection is crucial for diagnosing hematological diseases using microscopic images.
• Fully-supervised methods require extensive annotations, increasing costs and labor.
• Existing semi-supervised methods struggle with class and scale imbalances in blood cell images.

Purpose of the Study:

• To develop a versatile semi-supervised framework for blood cell detection that overcomes limitations of current methods.
• To improve the accuracy and efficiency of blood cell detection in microscopic images.
• To reduce the reliance on large-scale annotated datasets.

Main Methods:

• Proposes a Class and Scale Balanced semi-supervised framework for Blood Cell detection (CSBC).
• Employs a Dual-view High-quality Instance Selection (DHIS) strategy for reliable pseudo-label generation.
• Introduces a Class and Scale Balanced Resampling (CSBR) module and a Background-Aware Pasting Strategy (BAPS).

Main Results:

• CSBC demonstrates superior performance compared to ten state-of-the-art semi-supervised object detection methods.
• The framework shows consistent improvements under both Faster R-CNN and RetinaNet detectors.
• Experiments on BCCD and PBC datasets validate the effectiveness and generalizability of CSBC.

Conclusions:

• The proposed CSBC framework effectively addresses class and scale imbalances in blood cell detection.
• CSBC offers a robust and generalizable solution for semi-supervised blood cell detection.
• This approach significantly enhances diagnostic capabilities for hematological diseases through improved image analysis.