Validation study demonstrates 90.22% pixel accuracy in murine cervical tissue microscopy, indicating robust automated segmentation with limited training data.
Key Points
To develop an automated deep learning framework capable of accurately segmenting thin-section microscopy images using a minimally sized training dataset.
Constructed an image segmentation pipeline using a U-Net convolutional neural network architecture with a pretrained ResNet-34 encoder applied directly to total intensity images.
Trained and evaluated the model on 74 manually annotated murine uterine cervix sections (51 used for training, 23 held out for testing) across four anatomical classes: background, internal os, cervical tissue, and vaginal wall.
Achieved 90.22% pixel accuracy on the held-out test dataset using only 51 training images.
Successfully differentiated all four anatomical classes with minimal image pre-processing required prior to inference.