Lightweight Deep Learning Model for Classification of Normal and Abnormal Vasculature in Organoid Images.

Area of Science:

• Biomedical Engineering
• Computational Biology
• Cell Biology

Background:

• Human organoids model organ microenvironments for research.
• Assessing normal vasculature formation in organoids is crucial for experimental reliability.
• Current manual assessment methods are time-consuming and subjective.

Purpose of the Study:

• To develop a lightweight deep learning model for automated classification of normal and abnormal vasculature in vascular organoid images.
• To improve the efficiency and reproducibility of organoid vasculature assessment.

Main Methods:

• A modified EfficientNet model (replacing SiLU with ReLU, removing SE blocks) was used for image classification.
• The model was trained on vascular organoid images from co-culture experiments.
• Data augmentation and noise addition were employed to address class imbalance.

Main Results:

• The proposed Modified 3 models (B0, B1, B2) achieved high accuracy (0.90-1.00).
• Real-time inference speeds of 51.1, 36.0, and 32.4 FPS were recorded on a CPU.
• An average speed improvement of 70% was observed compared to original models.

Conclusions:

• The developed lightweight deep learning model enables efficient and automated vasculature assessment in organoids.
• The framework provides quantitative and reproducible analysis, enhancing the reliability of organoid-based research.
• The model's real-time processing capability supports high-throughput analysis.