End-to-End Deep Learning-Based Cells Detection in Microscopic Leucorrhea Images.

Area of Science:

• Gynecology
• Medical Imaging
• Artificial Intelligence

Background:

• Vaginitis is a common gynecologic condition affecting millions of women.
• Manual analysis of microscopic vaginal discharge images for vaginitis diagnosis is time-consuming and labor-intensive.
• Automating this process is crucial for early-stage diagnosis and efficient healthcare.

Purpose of the Study:

• To develop and evaluate an automated deep learning framework for detecting and identifying cells in microscopic leucorrhea images.
• To enable early-stage diagnosis of vaginitis through efficient image analysis.
• To achieve state-of-the-art performance in cell detection within microscopic vaginal discharge images.

Main Methods:

• An end-to-end deep learning framework utilizing the attention-based Detection with Transformers (DETR) architecture was proposed.
• Transfer learning was employed to accelerate network convergence and minimize annotation costs.
• A weighted sampler with on-the-fly data augmentation was integrated to handle class imbalance.
• The multi-head attention mechanism and bipartite matching loss of DETR were leveraged for real-time identification of overlapping cells.

Main Results:

• The proposed framework achieved a mean average precision (mAP) of 86.00%.
• Specific average precisions (AP) for detected cell types were: epithelium (96.76%), leukocyte (83.50%), pyocyte (74.20%), mildew (89.66%), and erythrocyte (88.80%).
• The average test time per image was approximately 72.3 ms, demonstrating real-time processing capabilities.

Conclusions:

• The developed deep learning-based cell detection framework offers a highly accurate and efficient solution for analyzing microscopic leucorrhea images.
• This automated approach significantly reduces diagnostic time and labor, representing a state-of-the-art method for vaginitis-related cell identification.
• The framework's performance, particularly in handling class imbalance and overlapping cells, shows great promise for clinical application in early vaginitis diagnosis.