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

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.

💡 Research Summary

The paper presents a streamlined, intensity‑based deep‑learning pipeline for segmenting tissue regions in Mueller matrix microscopy using only the total intensity element M₁₁. Conventional Mueller imaging analyses often require full matrix reconstruction, physical realizability checks, and complex preprocessing, which become problematic for thin sections where diagonal matrix elements approach unity and contrast is low. By focusing solely on M₁₁, the authors demonstrate that sufficient morphological contrast exists for reliable segmentation, dramatically simplifying the workflow.

A custom transmission‑mode Mueller polarimeter built on a commercial microscope captured 4 × 4 Mueller matrices of 70 murine uterine cervix sections (≈50 µm thick) across gestational days 0–18. From each matrix the M₁₁ intensity image was extracted, normalized, and resized to 512 × 512 pixels. To exploit pretrained ImageNet features, the single‑channel image was replicated to three channels and normalized with ImageNet mean and standard deviation.

For annotation, the authors developed a PyQt5 graphical tool that allows users to draw polygons or freehand masks for four semantic classes: background, general tissue, internal os, and vaginal wall. Masks were combined using a priority hierarchy (vaginal wall > os > tissue > background) to generate four‑class ground‑truth maps. The dataset was stratified by acquisition day and split into 49 training, 10 validation, and 11 test images, ensuring that temporal variations did not leak between sets.

The segmentation model is a U‑Net with a ResNet‑34 encoder pretrained on ImageNet‑1K. The encoder downsamples through five stages (64, 64, 128, 256, 512 channels) while the decoder restores spatial resolution via skip connections and bilinear up‑sampling. The output consists of four logits, one per class. Training employed a combined loss (0.5 × cross‑entropy + 0.5 × Dice) to balance class imbalance and boundary accuracy, the AdamW optimizer (lr = 1e‑4, weight decay = 1e‑5), batch size 8, and 50 epochs with mixed‑precision on an RTX 5000 GPU. Data augmentation included random horizontal/vertical flips, 90° rotations, and intensity jitter (±15%).

Results show rapid convergence: training accuracy 91.72 % and validation accuracy 89.66 % with a stable validation loss of 0.374. On the held‑out test set the model achieved 89.71 ± 5.51 % pixel accuracy and a mean Dice coefficient of 80.96 ± 13.37 %. Class‑wise Dice scores were 92.37 % (background), 88.63 % (tissue), 84.85 % (internal os), and 69.41 % ± 32.61 % (vaginal wall). The lower performance on the vaginal wall reflects genuine biological variability—its presence depends on sectioning depth—rather than a model flaw. Visual inspection confirmed accurate delineation of clear structures, correct identification of absent vaginal wall, and even detection of low‑contrast features missed by manual annotators. Failure cases were linked to tissue damage during preparation and poorly illuminated regions, highlighting sensitivity to image quality.

The discussion emphasizes three main contributions: (1) demonstrating that a single‑channel intensity image suffices for robust tissue segmentation in Mueller microscopy, eliminating the need for full matrix decomposition; (2) showing that ImageNet‑pretrained ResNet‑34 can transfer natural‑image features to polarimetric intensity patterns, enabling high performance with only 70 annotated samples—a stark contrast to the hundreds or thousands typically required for training from scratch; (3) providing an open‑source annotation GUI and trained model, facilitating reproducibility and extension to other tissues or imaging modalities. Limitations include susceptibility to domain shift (different tissue morphologies may require additional fine‑tuning), high variance in classes with intrinsic biological variability, and degraded performance under low signal‑to‑noise conditions. Future work could incorporate multi‑channel Mueller data, domain adaptation techniques, or more sophisticated noise‑robust training to broaden applicability.

Overall, this study offers a practical, low‑overhead solution for automated segmentation of Mueller microscopy data, paving the way for scalable quantitative polarimetric analyses in biomedical research where annotated datasets are scarce.