2SLOD-HCG: HCG Test Strip Concentration Prediction Network

Area of Science:

• Biomedical engineering
• Medical diagnostics
• Artificial intelligence in healthcare

Background:

• Human chorionic gonadotropin (HCG) is a critical biomarker for pregnancy detection.
• Current HCG test strip interpretation faces challenges including user error, AI limitations, and variable image quality.
• Accurate and accessible HCG detection is vital for timely diagnosis of pregnancy-related conditions.

Purpose of the Study:

• To develop a robust and accurate AI-based method for HCG test strip concentration detection.
• To overcome limitations of existing AI detection methods and improve reliability under diverse imaging conditions.
• To enhance the diagnostic capabilities for early pregnancy, multiple pregnancies, and ectopic pregnancies using automated test strip analysis.

Main Methods:

• Proposed 2SLOD-HCG, a novel network featuring an enhanced spatial pyramid pooling (SPP) module for multi-scale information integration.
• Incorporated an elastic variational cross-FPN with lightweight transformer blocks for improved global feature perception.
• Applied a SimAM attention mechanism to emphasize critical local features in test strip images.
• Created a dataset of 50,000 augmented test strip images under varied lighting and mobile photography conditions.

Main Results:

• The 2SLOD-HCG network demonstrated superior accuracy and robustness compared to YOLO-based baselines.
• The model excelled at detecting small, crucial color-developing regions on HCG test strips.
• Performance improvements were noted across various lighting conditions and mobile photography scenarios.

Conclusions:

• The 2SLOD-HCG network offers a significant advancement in automated HCG test strip analysis.
• This approach enhances the reliability and accuracy of AI-driven pregnancy diagnostics.
• The method shows promise for more accessible and precise early pregnancy evaluations.