Automated phenotyping of mammary gland tissues using computer vision systems

Area of Science:

• Dairy Science
• Biotechnology
• Animal Science

Background:

• Accurate quantification of mammary gland tissue composition is vital for understanding dairy heifer development and lactation potential.
• Traditional histological analysis is laborious and time-consuming, limiting large-scale studies.

Purpose of the Study:

• To develop and validate an automated deep learning-based phenotyping framework for classifying mammary gland tissues in Holstein heifers.
• To enable high-throughput, reproducible quantification of mammary tissue development across different ages.

Main Methods:

• Utilized deep learning semantic segmentation (U-Net and U-Net++) with ResNet34 encoders on manually annotated histological images.
• Trained models using focal loss and evaluated performance with Intersection over Union (IoU) metrics.
• Processed 130 images into 383,650 patches (448x448 pixels) for model training and validation.

Main Results:

• The U-Net++ model without an attention module achieved the highest IoU values, demonstrating high prediction accuracy.
• Mean average percentage error was below 5% for most tissue-time points, with ductal tissue showing the highest accuracy.
• Adipose tissue exhibited dynamic growth, increasing significantly from 10 to 39 weeks of age.

Conclusions:

• The developed automated framework offers a scalable and reproducible solution for histological phenotyping of mammary gland tissue.
• This approach facilitates high-throughput quantification of mammary tissue development, overcoming limitations of manual methods.
• The study provides a valuable tool for dairy science research, bridging detailed tissue characterization with large-scale phenotyping needs.