Informatics, digital & computational pathology

Computational pathology

Computational pathology fundamentals & applications

Last author update: 5 December 2022
Last staff update: 5 December 2022

Copyright: 2019-2023,, Inc.

PubMed Search: Computational pathology

Yomna Amer, M.B.B.Ch.
Anil Parwani, M.D., Ph.D., M.B.A.
Page views in 2022: 884
Page views in 2023 to date: 815
Cite this page: Amer Y, Parwani A. Computational pathology fundamentals & applications. website. Accessed September 29th, 2023.
Definition / general
  • Computational pathology is an approach to diagnosis that integrates data from numerous sources (pathology, molecular, radiology and clinical), using mathematical models to generate clinically applicable knowledge; it facilitates the best possible medical decision making by enhancing diagnostic accuracy and reduces cost by improving lab efficiency (Arch Pathol Lab Med 2014;138:1133)
Essential features
  • Artificial intelligence (AI): a branch of computer science dealing with the simulation of human intelligent behavior in computers; the main goal of pathology AI is to analyze digital slide images via image analysis and machine learning (Lab Invest 2021;101:412)
  • Machine learning (ML): a branch of artificial intelligence where software learns to perform mathematical models by being exposed to representative data in order to generate expectations or decisions (Lab Invest 2021;101:412)
  • Deep learning (DL): a class of machine learning algorithms that is composed of complex neural networks where the software trains itself to perform tasks based on unstructured or unlabeled data (Lab Invest 2021;101:412)
  • Digital pathology (virtual microscopy): utilization of microscopy and digital technology to obtain images using whole slide image scanners, analyze them using image viewers (usually computer monitors) and infer digital information (Nat Rev Nephrol 2020;16:669)
  • Whole slide image (WSI) (virtual slide): a high resolution copy of a glass slide that is created by a slide scanner and can be perceived on a computer screen (Nat Rev Nephrol 2020;16:669)
  • Segmentation (annotation): one of the principal tasks of machine vision
    • Manual segmentation (annotation) is performed by experts to define the boundaries of the object (region of interest)
    • Automatic segmentation is achieved by the deep learning model that is trained to distinguish the boundaries of the object (Nat Rev Nephrol 2020;16:669)
  • Supervised machine learning: analysis of labeled datasets to train or supervise algorithms to accurately predict outcomes (BMC Med Inform Decis Mak 2019;19:281)
  • Unsupervised machine learning: algorithms that analyze unlabeled data sets to discover hidden patterns in data without the need for human intervention (hence, they are unsupervised) (BMC Med Inform Decis Mak 2019;19:281)
  • Pathomics: the wide variety of data that results from the analysis of digitized histopathology images and generates detailed features of the entire tissue sample in whole slide images (Chin J Cancer Res 2021;33:563)
Examples of uses of computational pathology
  • Applications in colorectal cancer:
    • Developing multiple deep learning algorithms, which can precisely categorize whole slide images of 5 types of colorectal polyps: hyperplastic, sessile serrated, traditional serrated, tubular and tubulovillous / villous polyps (J Pathol Inform 2017;8:30)
    • Predicting colorectal cancer outcomes based on tissue microarray samples by a combination of convolutional neural networks and recurrent neural network architectures (Sci Rep 2018;8:3395)
  • Applications in cytopathology:
    • Classification of urine cytology whole slide images based on the Paris System by using morphometric algorithm and semantic segmentation network (Nat Rev Genet 2015;16:321)
  • Application in prostate cancer:
    • Machine learning techniques are utilized to identify either the individual genes or groups of them to predict the clinical outcome by using genomic based risk prediction models and molecular profiling (Nat Rev Genet 2015;16:321)
  • Applications in breast cancer:
    • Effectively analyze mitosis in a histopathological tissue sample for breast cancer grading using a convolutional neural network (CNN) and semantic segmentation (Front Bioeng Biotechnol 2019;7:145)
    • Classification of breast cancer pathology images for both benign and malignant categories via convolutional neural network (Front Bioeng Biotechnol 2019;7:145)
Diagrams / tables

Traditional image analysis versus computational pathology
Traditional image analysis
(J Pathol 2019;249:286)
Deep learning powered computational pathology
(J Pathol Inform 2019;10:9)
Typical tasks Detection of a morphological pattern (Lab Invest 2021;101:412) Integration of all aspects of clinical workflow for more accurate diagnosis, prognosis and personalized treatment (Lab Invest 2021;101:412)
Parameter tuning Image features / parameters are
manually tuned
Algorithm learns and extracts a large number of features automatically
Typical algorithm testing Often on a few regions of the slide Usually whole slide
Computer unit best suited for task CPU (central processing unit) GPU (graphics processing unit)
Number of training images required Depending on application, may be low Usually remarkably high
Challenges facing computational pathology
  • Standardization and normalization - inaccurate raw data and unreliable results from low quality of slide preparation, which includes
    • Folding in tissue section
    • Existence of air bubble
    • Different sets of illumination
    • Intensity discrepancy
    • Average color
    • Boundary intensity during scanning (Lab Invest 2021;101:412)
  • Hardware limitations - requirements for precise application of computational pathology include
  • Ethical issues:
    • Transfer of a massive amount of health data among clinics, laboratories and data banks raise security vulnerability (Lab Invest 2021;101:412, J Med Ethics 2022;48:278)
    • Inadequate transparency and inaccessibility to significant patients' data set characteristics, including race and ethnicity, which affect generalizability of the of the model (Nat Med 2020;26:1364)
Challenges facing traditional pathology
  • Some pathological disease classifications are highly tedious, as several categories may require pathologists to reach the decision (Lab Invest 2021;101:412)
  • Inability to detect specific criteria may lead to errors in diagnosis (Lab Invest 2021;101:412)
  • The strength of interobserver agreement is often average or even poor for some lesions, despite carefully explained classifications (Lab Invest 2021;101:412)
Board review style question #1
Which of the following is feature of a deep learning (DL) based algorithm that is not typical of traditional image analysis algorithms?

  1. It can segment and classify objects in H&E images
  2. It uses handcrafted image features, such as cell shape, to classify objects
  3. It learns by extracting important image features automatically from examples
  4. It can measure staining intensity of labeled objects from IHC images
Board review style answer #1
C. The DL based algorithm learns by extracting important image features automatically from examples. C is the correct answer as it is the primary feature of deep learning based algorithms but not of traditional image analysis algorithms. Answer A is a task that can be accomplished by both types of methods. Answers B and D are commonly performed by traditional image analysis algorithms.

Comment Here

Reference: Computational pathology fundamentals & applications
Board review style question #2
Which of the following statements about supervised deep learning is true?

  1. Images used for training must be representative of the images that the algorithm is designed to be applied to
  2. To reduce bias, images used to train algorithms should be carefully selected to contain minimal artifacts
  3. Only a small number of images is typically required for training
  4. Trained algorithms should be tested in datasets that were also used for training
Board review style answer #2
A. Images used for training must be representative of the images that the algorithm is designed to be applied to. Answer A is the only true statement. Answer B is false because images should contain artifacts represented in the images that the algorithm is designed to be applied to. Answer C is false because deep learning typically requires a large number of images to train. Answer D is false because trained algorithms need to be tested in an entirely independent dataset.

Comment Here

Reference: Computational pathology fundamentals & applications
Back to top
Image 01 Image 02