Microscopy image segmentation using a fine-tuned machine learning model with limited training dataset.

Abstract

Manual annotation of microscopic images of thin tissue sections remains time-consuming step limiting the throughput of diagnostic analysis. We present a novel automated approach using the total intensity image as an input to a U-Net architecture of Convolutional Neural Network with a pretrained ResNet-34 encoder for image segmentation. The network was trained on limited-size annotated dataset to distinguish four classes in the images of thin sections of murine uterine cervix: background, internal os, cervical tissue, and vaginal wall. With only 74 manually annotated images in total (51 used for model training), the model achieved 90.22% pixel accuracy on the held-out test dataset. This framework requires minimal data pre-processing and is readily extensible to other tissue types, with publicly available graphical annotation tools for practical deployment.