Large scale digital prostate pathology image analysis combining feature extraction and deep neural network

Naiyun Zhou, Andriy Fedorov, Fiona M. Fennessy, Ron Kikinis, Yi GaoarXiv 2017 · DOI · arXiv

Abstract

Histopathological assessments, including surgical resection and core needle biopsy, are the standard procedures in the diagnosis of prostate cancer. Current interpretation of histopathology images includes the determination of the tumor area, Gleason grading, and identification of certain prognosis-critical features. Such a process is not only tedious but also prone to intra/inter-observer variabilities. This work presents an analysis pipeline that includes localization of the cancer region, grading, area ratio of different Gleason grades, and cytological/architectural feature extraction. The proposed algorithm combines human-engineered feature extraction with features learned by a deep neural network. The entire pipeline is implemented to directly operate on whole slide images (WSIs) produced by digital scanners. The algorithm was tested on 368 whole slide images from the TCGA dataset and achieved an overall accuracy of 75% in differentiating Gleason 3+4 from 4+3 slides.

Introduction

Prostate cancer (PCa) diagnosis relies heavily on the Gleason grading system, which evaluates the architectural patterns of cancer cells. However, manual grading is tedious and suffers from high inter-observer variability—agreement among pathologists can be as low as 38%. The challenge is particularly acute for "Gleason 7" cases, which are split into 3+4 (mostly grade 3) and 4+3 (mostly grade 4). This distinction is vital because the 10-year survival rate for 3+4 is significantly higher (92.1%) than for 4+3 (76.5%). This research aims to create a reproducible, automated pipeline that can operate directly on high-resolution Whole Slide Images (WSIs) to assist clinicians in making these high-stakes distinctions.

Methods

The proposed multi-stage analysis pipeline integrates traditional image processing with deep learning to handle the complexity of WSIs:

  1. Tumor Region Extraction: To focus on relevant tissue, K-means clustering in the Lab* color space is used to localize gland regions and exclude stroma.
  2. Optimized Color Decomposition: A novel color decomposition method using regularized minimization adaptively extracts the hematoxylin channel, ensuring robustness across different scanners and stain conditions.
  3. Deep Learning Architecture: An 8-layer Convolutional Neural Network (CNN) was developed, consisting of 6 convolutional layers and 2 fully connected layers. The network was trained on 256x256 patches using a "label-free" strategy, where "pure" slides (Gleason 3+3 and 4+4) provided the fundamental features without requiring pixel-level annotations.
The 8-layer CNN architecture used for feature learning.
  1. Prognosis-Significant Feature Detection: Specialized algorithms detect prominent nucleoli (using EM analysis) and cribriform patterns (using graph theory and geometric metrics), providing interpretable clinical indicators that standard CNNs might overlook.

Results and Discussion

The study utilized 368 digitized pathology slides from The Cancer Genome Atlas (TCGA) Prostate Adenocarcinoma dataset. The pipeline achieved a 75% overall accuracy in differentiating Gleason 3+4 from 4+3 slides.

Representative whole slide images and patches for Gleason 3+4 and 4+3.

The hybrid approach of combining CNNs with human-engineered features provides a more interpretable and clinically relevant description of the disease compared to "black box" models. By using slide-level labels for training, the method demonstrates a path toward scaling AI in pathology where expert annotation time is limited.

While successful, the current framework is constrained to binary classification (Grade 3 vs. 4) and lacks local ground truth for precise localization validation. Future work will focus on expanding the model to handle benign tissue and Grade 5 patterns, as well as refining the quantification of intermixed Gleason patterns to further improve clinical utility.

Summary updated

Abstract

Histopathological assessments, including surgical resection and core needle biopsy, are the standard procedures in the diagnosis of the prostate cancer. Current interpretation of the histopathology images includes the determination of the tumor area, Gleason grading, and identification of certain prognosis-critical features. Such a process is not only tedious, but also prune to intra/inter-observe variabilities. Recently, FDA cleared the marketing of the first whole slide imaging system for digital pathology. This opens a new era for the computer aided prostate image analysis and feature extraction based on the digital histopathology images. In this work, we present an analysis pipeline that includes localization of the cancer region, grading, area ratio of different Gleason grades, and cytological/architectural feature extraction. The proposed algorithm combines the human engineered feature extraction as well as those learned by the deep neural network. Moreover, the entire pipeline is implemented to directly operate on the whole slide images produced by the digital scanners and is therefore potentially easy to translate into clinical practices. The algorithm is tested on 368 whole slide images from the TCGA data set and achieves an overall accuracy of 75% in differentiating Gleason 3+4 with 4+3 slides.
BibTeX
@misc{zhou2017large,
  author = {Naiyun Zhou and Andriy Fedorov and Fiona M. Fennessy and Ron Kikinis and Yi Gao},
  title = {{Large scale digital prostate pathology image analysis combining feature extraction and deep neural network}},
  year = {2017},
  eprint = {1705.02678},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1705.02678},
}
Figure