Morphological components detection for super-depth-of-field bio-micrograph based on deep learning

Area of Science:

• Medical Imaging
• Computer Vision
• Biotechnology

Background:

• Clinical routine examination demands are increasing, necessitating higher efficiency and accuracy.
• Automatic classification and localization of cells in super depth of Field (SDoF) microscopic images present significant challenges.

Purpose of the Study:

• To develop and evaluate an advanced object detection algorithm for cells in SDoF micrographs.
• To improve the accuracy and efficiency of cell detection in clinical diagnostic applications.

Main Methods:

• An object detection algorithm based on the Retinanet model was advanced for cell detection in SDoF micrographs.
• The algorithm was experimentally validated using leucorrhea and fecal samples.

Main Results:

• The proposed Retinanet-based algorithm demonstrated significant improvements in mean average precision (mAP) compared to mainstream methods.
• mAP indexes reached 83.1% for leucorrhea samples and 88.1% for fecal samples, an average increase of 10%.

Conclusions:

• The developed object detection model effectively addresses the challenges of cell classification and localization in SDoF systems.
• This model is applicable to feces and leucorrhea detection equipment, promising substantial improvements in diagnostic efficiency and accuracy.