Source-linked AI summary
Epithelium segmentation using deep learning in H&E-stained prostate specimens with immunohistochemistry as reference standard
Wouter Bulten, Péter Bándi, Jeffrey Hoven, Rob van de Loo, Johannes Lotz, Nick Weiss, Jeroen van der Laak, Bram van Ginneken, Christina Hulsbergen-van de Kaa, Geert Litjens
TL;DR
The paper addresses noisy, coarse epithelial annotations that limit automated prostate-cancer analysis. It trains an IHC-based U-Net to generate precise epithelial masks, transfers them to H&E, and trains a final H&E segmentation network. The system produces cell-level segmentation of intact glands and individual tumor cells, with performance reported across hold-out and external data.
Problem
Coarse prostate-cancer annotations contain non-relevant tissue, while high-grade tumors and individual cells are difficult to annotate precisely.
Method
An IHC U-Net generates epithelial masks that are registered to H&E slides and used to train a final H&E segmentation network.
Results
The system produces cell-level segmentations of intact glands and individual tumor epithelial cells, and achieves a highest reported hold-out F1-score of 0.915.
Takeaways & Limitations
The segmentation can precisely outline potential cancer regions and combine with coarse tumor annotations to produce detailed prostate-cancer annotations.
Takeaways & Limitations
IHC-derived labels remain vulnerable to stain variability, scan and tissue artefacts, stained debris, and missed epithelium when staining is light or absent.
Abstract
from arXiv · showhide
Prostate cancer (PCa) is graded by pathologists by examining the architectural pattern of cancerous epithelial tissue on hematoxylin and eosin (H&E) stained slides. Given the importance of gland morphology, automatically differentiating between glandular epithelial tissue and other tissues is an important prerequisite for the development of automated methods for detecting PCa. We propose a new method, using deep learning, for automatically segmenting epithelial tissue in digitized prostatectomy slides. We employed immunohistochemistry (IHC) to render the ground truth less subjective and more precise compared to manual outlining on H&E slides, especially in areas with high-grade and poorly differentiated PCa. Our dataset consisted of 102 tissue blocks, including both low and high grade PCa. From each block a single new section was cut, stained with H&E, scanned, restained using P63 and CK8/18 to highlight the epithelial structure, and scanned again. The H&E slides were co-registered to the IHC slides. On a subset of the IHC slides we applied color deconvolution, corrected stain errors manually, and trained a U-Net to perform segmentation of epithelial structures. Whole-slide segmentation masks generated by the IHC U-Net were used to train a second U-Net on H&E. Our system makes precise cell-level segmentations and segments both intact glands as well as individual (tumor) epithelial cells. We achieved an F1-score of 0.895 on a hold-out test set and 0.827 on an external reference set from a different center. We envision this segmentation as being the first part of a fully automated prostate cancer detection and grading pipeline.
Introduction
Prostate cancer grading depends on epithelial architecture, but pathologists face laborious, coarse annotation tasks, especially in poorly differentiated tumors. The paper proposes pixel-level epithelial segmentation using deep learning and IHC-derived training labels.
- 1.1 million new diagnoses occur annually, making prostate cancer the most common cancer in men in developed countries.
- High-grade tumors may lose glandular structure and appear as strands of individual cells, complicating epithelial annotation.
- Gleason grading assigns architectural patterns from 1 to 5, with higher numbers indicating lower differentiation and typically worse prognosis.
- Deep-learning diagnosis is motivated by the time-consuming analysis of multifocal disease across numerous surgical-specimen sections.
- Pathologist annotations are often coarse and include non-relevant tissue, adding noise that limits deep-learning methods.
- The proposed system divides digitized tissue pixel by pixel into epithelial and other tissue, using IHC segmentation transferred to H&E training.
Related work
Prior prostate epithelium-segmentation studies used classical or deep-learning methods but often relied on small datasets or limited Gleason grades. This work emphasizes coverage across grades and gland morphologies.
- Existing methods include support vector machines using handcrafted H&E features and Bayesian classifiers relying on gland lumina.
- Deep-learning methods have outperformed handcrafted features on several digital-pathology segmentation tasks.
- A prior pilot study using 30 prostatectomy slides found U-Net effective, but reference-standard errors and few high-grade cases limited applicability.
- Most existing prostate epithelium-segmentation studies use small datasets or focus on only some occurring grades.
- The present study includes all Gleason grades and gland morphologies rather than excluding difficult patterns.
Materials
The dataset contains 102 prostatectomy tissue blocks with paired H&E and IHC sections, stratified train/test division, graded tumor regions, and an external annotated reference set.
- The cohort comprised 102 radical-prostatectomy patients selected according to reported Gleason grades.
- Each block yielded a newly cut H&E section that was scanned, destained, restained with IHC, and scanned again at 20x magnification.
- CK8/18 marked benign and malignant glandular epithelium, while P63 marked the basal layer normally present in benign glands.
- The 102 slide pairs were randomly split into 62 training and 40 hold-out test slides, stratified by Gleason grade group.
- Test regions included tumor and benign epithelium and were graded independently by an experienced uropathologist.
- The external reference set contained 224 tiles from 20 H&E prostatectomy whole-slide images with Gleason grades 3 and 4.
Methods
The method uses IHC to generate epithelial masks, transfers them through nonlinear registration to H&E, and trains a final H&E U-Net. Manual correction and artifact comparisons address errors in stain-derived labels.
- Pipeline: The two-step pipeline trains an IHC segmentation network first, then uses its registered outputs as training targets for the H&E network.
- IHC labels: Color deconvolution and thresholding convert P63/CK8-18 staining into binary epithelial masks, with morphological cleanup of small errors.
- IHC labels: Manually corrected regions were used to train the IHC network, while three trained non-experts updated masks in sampled test regions.
- IHC labels: Data augmentation included flips, rotations, noise, blurring, and color changes to improve generalization across stain variation.
- H&E network: A six-level H&E U-Net was used to reduce misclassified corpora amylacea, compared with a five-level IHC network.
- H&E network: Pixel-class weighting encouraged detection of small epithelial areas, including those associated with Gleason 5 tumors.
Results
The IHC network achieved strong but stain-sensitive epithelial segmentation, while the H&E network maintained high performance on hold-out and external data, including challenging tumor regions.
- IHC segmentation: 0.915 overall F1-score was achieved by the IHC network, although performance varied with stain quality and intensity.Scores ranged from 0.352 to 0.980, and lower stain intensity reduced performance.
- IHC segmentation: 0.909 to 0.915 F1-score improvement resulted from correcting color deconvolution masks before training.Uncorrected training produced more errors in stained non-epithelial tissue such as corpora amylacea and concretions.
- H&E segmentation: 0.893 overall F1-score was achieved by the H&E network, with higher performance on benign tissue than tumorous areas.Benign tissue reached F1 0.907 versus 0.876 for tumorous areas; Gleason grade group 5 reached F1 0.819.
- H&E segmentation: 0.661 versus 0.352 minimum performance favored the H&E network over the IHC network, whose overall score was otherwise comparable.Outliers present in IHC results were absent from H&E results.
- H&E segmentation: 0.893 versus 0.878 F1-score resulted when IHC-network outputs replaced raw color deconvolution masks as H&E training data.The comparison shows improved H&E segmentation from the IHC-generated training masks.
- External validation: 0.835 F1-score was achieved on the external set despite differences in staining and image resolution.Its Jaccard score was 0.735, compared with 0.595 for the original method.
Discussion
The deep learning system segments prostate epithelial tissue at cell level on H&E slides, covering intact glands and individual tumor cells. IHC-derived training labels support performance across Gleason grades, while stain variability, magnification, and H&E ambiguity remain important constraints.
- The system produces cell-level epithelial segmentations for both intact glands and individual tumor epithelial cells.
- IHC-derived labels reduce manual annotation burden and help capture small epithelial instances that human annotators might miss, especially in Gleason 5 cancer.
- 0.819 F1-score was achieved on Gleason grade 5, the lowest score across the assessed patterns but the first benchmark reported for these grades.
- 0.893 versus 0.878 F1 score was obtained when IHC network outputs trained the H&E network rather than using non-corrected data directly.
- High magnification supports individual-cell segmentation, whereas lower magnification may better segment intact glands and reduce artefacts.
- The H&E network approaches IHC performance, but further gains are limited because some cell classes cannot be assessed from H&E alone, particularly during active inflammation.