Artificial Intelligence Model Training for Blood Cell Region Segmentation
Universidade de São Paulo
Jan 2022 - Jun 2023 (1 year 6 months)
This work involved training and implementing an AI model for segmenting and classifying blood cell regions, including basophils, eosinophils, lymphocytes, monocytes, and neutrophils. Using the U-Net architecture, designed for biomedical image segmentation, the neural network achieved 99.34% accuracy and 1.03% loss after 30 training epochs on a dataset of 1145 images. The Sørensen-Dice coefficient reached 97.16%, reflecting high segmentation quality. Data augmentation enhanced the variety of training images. The model delivered highly accurate segmentations, with minimal discrepancies for complex cell types, ensuring reliable and precise results.
