Intensity-based Segmentation of Tissue Images Using a U-Net with a Pretrained ResNet-34 Encoder: Application to Mueller Microscopy

TL;DR

This work addresses automated segmentation of Mueller-matrix tissue images, where thin sections yield weak polarimetric contrast, by using only the total intensity image $M_{11}$ as input to a U-Net with a pretrained ResNet-34 encoder. The method leverages transfer learning from ImageNet to achieve accurate, pixel-wise segmentation on a small dataset (70 annotated sections), reporting a held-out test pixel accuracy of 89.71% and a mean tissue Dice coefficient of 80.96%. Key contributions include demonstrating that single-channel intensity patterns can transfer learned features to polarimetric data, and providing an accessible annotation and deployment pipeline. The approach offers minimal preprocessing, extensibility to other tissues and modalities, and practical deployment potential in resource-limited settings.

Abstract

Manual annotation of the images of thin tissue sections remains a time-consuming step in Mueller microscopy and limits its scalability. We present a novel automated approach using only the total intensity M11 element of the Mueller matrix as an input to a U-Net architecture with a pretrained ResNet-34 encoder. The network was trained to distinguish four classes in the images of murine uterine cervix sections: background, internal os, cervical tissue, and vaginal wall. With only 70 cervical tissue sections, the model achieved 89.71% pixel accuracy and 80.96% mean tissue Dice coefficient on the held-out test dataset. Transfer learning from ImageNet enables accurate segmentation despite limited size of training dataset typical of specialized biomedical imaging. This intensity-based framework requires minimal preprocessing and is readily extensible to other imaging modalities and tissue types, with publicly available graphical annotation tools for practical deployment.

Figures (7)

• Figure 1: Images of murine uterine cervix thin section (Day 18 of the gestation). Left: MM images. Color bar: $[0,1]$ (diagonal elements), $[-0.1,0.1]$(off-diagonal elements). Right: Derived maps of polarimetric parameters using Lu-Chipman decomposition LC: depolarization $\Delta$ (dimensionless), diattenuation $D$ (dimensionless), linear retardance $LR$ (rad), and optical axis azimuth $\psi$ (rad).
• Figure 2: Screenshot of the custom Tissue Annotation Tool GUI. The interface features an interactive canvas for visualizing the normalized $M_{11}$ intensity map. The control panel enables users to switch between vertex-based Polygon and Freehand drawing modes, adjust mask layer opacity, and define the semantic classes (e.g., internal os, vaginal wall)
• Figure 3: Representative images of cervical tissue sections. Left to Right: M$_{11}$ intensity image, tissue mask, OS mask, and vaginal wall mask. Top: Clear OS and vaginal wall structures. Middle: Absent vaginal walls. Bottom: Irregular internal vaginal wall shapes and ambiguous tissue borders.
• Figure 4: U-Net with a ResNet-34 encoder for M$_{11}$ tissue image segmentation. The encoder downsamples the input (purple arrows) with internal cross-layer connections (blue arrows). Skip connections (gray dashed arrows) pass high-resolution features to the decoder, which upsamples the features (orange arrows) with internal cross-layer connections (green arrows). Spatial resolution decreases from $256\times256$ to $16\times16$ at the bottleneck and increases to $512\times512$ at the output.
• Figure 5: Training metrics over 50 epochs. Left: combined loss showing steady convergence. Right: pixel accuracy demonstrating robust generalization.