Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning Melanocytic Cell Masks from Adjacent Stained Tissue

Mikio Tada, Ursula E. Lang, Iwei Yeh, Elizabeth S. Keiser, Maria L. Wei, Michael J. Keiser

TL;DR

This work tackles the variability in melanoma diagnosis by proposing a pixel-level melanocytic cell segmentation method that learns from paired H&E and adjacent Melan-A IHC tissue sections. A automated pipeline aligns H&E with IHC, uses color deconvolution and dilation to generate ground-truth masks, and trains a UNet with a combined Focal and Tversky loss on 512×512 tiles, achieving a mean IOU of $0.64$ on held-out slides. The approach emphasizes robust ground-truth construction through IHC-cytoplasm intersection and image registration, revealing higher performance in melanoma regions and highlighting the importance of artifact removal. This methodology enables in silico IHC labeling without tissue destaining, with potential generalization to other tissue types and antibodies, thus reducing the burden of gigapixel manual annotation in pathology workflows.

Abstract

Melanoma is one of the most aggressive forms of skin cancer, causing a large proportion of skin cancer deaths. However, melanoma diagnoses by pathologists shows low interrater reliability. As melanoma is a cancer of the melanocyte, there is a clear need to develop a melanocytic cell segmentation tool that is agnostic to pathologist variability and automates pixel-level annotation. Gigapixel-level pathologist labeling, however, is impractical. Herein, we propose a means to train deep neural networks for melanocytic cell segmentation from hematoxylin and eosin (H&E) stained sections and paired immunohistochemistry (IHC) of adjacent tissue sections, achieving a mean IOU of 0.64 despite imperfect ground-truth labels.

Learning Melanocytic Cell Masks from Adjacent Stained Tissue

TL;DR

This work tackles the variability in melanoma diagnosis by proposing a pixel-level melanocytic cell segmentation method that learns from paired H&E and adjacent Melan-A IHC tissue sections. A automated pipeline aligns H&E with IHC, uses color deconvolution and dilation to generate ground-truth masks, and trains a UNet with a combined Focal and Tversky loss on 512×512 tiles, achieving a mean IOU of on held-out slides. The approach emphasizes robust ground-truth construction through IHC-cytoplasm intersection and image registration, revealing higher performance in melanoma regions and highlighting the importance of artifact removal. This methodology enables in silico IHC labeling without tissue destaining, with potential generalization to other tissue types and antibodies, thus reducing the burden of gigapixel manual annotation in pathology workflows.

Abstract

Melanoma is one of the most aggressive forms of skin cancer, causing a large proportion of skin cancer deaths. However, melanoma diagnoses by pathologists shows low interrater reliability. As melanoma is a cancer of the melanocyte, there is a clear need to develop a melanocytic cell segmentation tool that is agnostic to pathologist variability and automates pixel-level annotation. Gigapixel-level pathologist labeling, however, is impractical. Herein, we propose a means to train deep neural networks for melanocytic cell segmentation from hematoxylin and eosin (H&E) stained sections and paired immunohistochemistry (IHC) of adjacent tissue sections, achieving a mean IOU of 0.64 despite imperfect ground-truth labels.
Paper Structure (9 sections, 1 equation, 3 figures)

This paper contains 9 sections, 1 equation, 3 figures.

Table of Contents

  1. Introduction
  2. Related Studies
  3. Method
  4. Dataset
  5. Automated melanocytic cell label generation
  6. Model training and evaluation
  7. Results
  8. Conclusion
  9. Acknowledgements

Figures (3)

  • Figure 1: Overview of an automated method to generate ground truth label masks using H&E and IHC stained tissues. The workflow begins (at left) with H&E and IHC WSIs and ends with a binary melanocyte ground-truth mask. The operations are done at the section-level, the model takes 512x512 tiles as input.
  • Figure 2: (a) Effect of image registration to alignment between H&E and IHC. (b) An example of spurious blue ink. (c) Demonstration of expanding the initial nuclei segmentation to cytoplasm.
  • Figure 3: (a) From left to right, examples of pairs of H&E and IHC tiles, the ground truth (green), the raw prediction, and the agreement between ground truth and the class prediction, for benign nevus and melanoma. (b) Distribution of mean IOUs based on the entire test set, benign nevus alone, and melanoma alone.