R-005 · Master's Research
Cell Segmentation U-Net Model
Cell classification and segmentation model trained on a Kaggle dataset.
2025 · Python · PyTorch · Machine vision · Data augmentation
Overview
This project was initially developed to test cell segmentation and verify a pipeline which would later be used in my self-driving lab work.
Design
Like many methods in the literature, I used a pre-trained U-Net before training on another 4000 images. The images were a combination of regular and augmented images, since the original Kaggle dataset was sparse. 65% of the 4000 images were augmented with rotations, flips and Gaussian blur. After the model generated a black and white image mask, OpenCV was used to count the number of cells within the mask.
Results
After 100 epochs, the model achieved a pixel accuracy of 95.83% on test data. In total, the model overpredicted the cell count by 2.27%. This model laid the foundation for developing an optical scanner specifically for the purpose of cell counting. While the model here was not state of the art, it proved that cells could be counted with reasonable accuracy using optical methods. Future work can compare this U-Net model to other models in the literature, or to other cell types and densities.