Table of Contents
Definition / general | Essential features | Terminology | Diagrams / tables | Basic considerations for supervised deep learning algorithm development | Examples of uses of computational pathology | Board review style question #1 | Board review style answer #1 | Board review style question #2 | Board review style answer #2Cite this page: Kozlowski C. Computational pathology fundamentals & applications. PathologyOutlines.com website. https://www.pathologyoutlines.com/topic/informaticscomppathfund.html. Accessed January 24th, 2021.
Definition / general
- A branch of pathology that involves extraction of information from digitized pathology images in combination with their associated metadata, typically using artificial intelligence methods such as deep learning
Essential features
- Computational pathology is the analysis of digitized pathology images with associated metadata, typically using artificial intelligence (AI) methods
- Deep learning is a type of AI method, commonly used in computational pathology, that is able to "learn" how to perform tasks based on examples
- Training a typical supervised deep learning algorithm in computational pathology involves large amounts of labeled training data
Terminology
- Artificial intelligence (AI): branch of computer science dealing with simulation of intelligent behavior in computers
- Cloud computing: using a network of remote servers to store, manage and process data
- Convolutional neural network (CNN): type of deep neural network particularly designed for images
- Deep learning: complex neural networks where software trains itself to perform tasks
- Machine learning: branch of AI where software learns to perform tasks by being exposed to representative data
- Supervised machine learning: training based on labels associated with a data point (i.e. ground truth)
- Unsupervised machine learning: training based on natural divisions in a data set without the need for a ground truth, e.g. clustering methods
- Reference: J Pathol 2019;249:286
Diagrams / tables
Traditional image analysis versus computational pathology
“Traditional” image analysis (J Pathol Inform 2019;10:9) |
Deep learning powered computational pathology (J Pathol 2019;249:286) |
|
Typical tasks | Automation of repetitive tasks, e.g.:
|
Image features correlated to patient meta data in order to, e.g.:
|
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 very high |
Basic considerations for supervised deep learning algorithm development
- Obtaining ground truth data
- Patient outcome data
- A field from the pathology report or laboratory information system
- A quantitative score assigned to the case
- Manually provided by a pathologist
- Considerations in acquisition
- Streamlined workflows
- Single common annotation tool
- Tradeoff between quantity and accuracy
- Good practices
- Training images must be representative of the images algorithm is designed to be applied to
- Use a wide variety of data sources
- Use consistent pre-imaging steps
- Apply manual or automated image quality control processes
- Use larger and more representative training sets
- Calibrate algorithms for each lab prior to being used for clinical work
- Apply image preprocessing strategies such as color normalization
- Data augmentation to artificially add variation and increase (or balance) the training data
- Test developed models using a variety of test and validation sets to avoid overfitting
Examples of uses of computational pathology
- "Predicting cancer outcomes from histology and genomics using convolutional networks" (Proc Natl Acad Sci U S A 2018;115:E2970)
- Used information from histology images and genomic biomarkers to predict survival in patients with glioma
- Combined convolutional neural network with survival models to predict time-to-event data
- Heatmap visualization to show image features related to prognosis
-
"Automated Gleason grading of prostate cancer tissue microarrays via deep learning" (Sci Rep 2018;8:12054)
- Used deep learning to train an algorithm to grade prostate cancer
- Used annotated tissue microarray images of prostate cancer to train algorithm
- Used transfer learning (network trained on another task and added their own images) so that they could build an algorithm with only 641 patients, a relatively small sample size
- Used "class activation mapping" to confirm that the model was focusing on important image features
Board review style question #1
- Which of the following statements could be true for a deep learning based algorithm but typically not for a traditional image analysis algorithm?
- The algorithm can segment and classify objects in H&E images
- The algorithm uses handcrafted image features, such as cell shape, to classify objects
- The algorithm "learns" by extracting important image features automatically from examples
- The algorithm can measure staining intensity of labeled objects from IHC images
Board review style answer #1
C. The algorithm "learns" by extracting important image features automatically from examples
Explanation: C is the correct answer as it is the primary feature of deep learning based algorithms but not of traditional image analysis algorithms. A is a task that can be accomplished by both types of methods. B and D are commonly performed by traditional image analysis algorithms.
Reference: Computational pathology fundamentals
Comment Here
Explanation: C is the correct answer as it is the primary feature of deep learning based algorithms but not of traditional image analysis algorithms. A is a task that can be accomplished by both types of methods. B and D are commonly performed by traditional image analysis algorithms.
Reference: Computational pathology fundamentals
Comment Here
Board review style question #2
- Which of the following statements about supervised deep learning is true?
- Images used for training must be representative of the images the algorithm is designed to be applied to
- To reduce bias, images used to train algorithms should be carefully selected to contain minimal artifacts
- Because deep learning is very powerful, only a small number of images is typically required for training
- 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 the algorithm is designed to be applied to
Explanation: A is the only true statement. B is false because images should contain artifacts represented in the images the algorithm is designed to be applied to. C is false because deep learning typically requires a large number of images to train. D is false because trained algorithms need to be tested in an entirely independent dataset.
Reference: Computational pathology fundamentals
Comment Here
Explanation: A is the only true statement. B is false because images should contain artifacts represented in the images the algorithm is designed to be applied to. C is false because deep learning typically requires a large number of images to train. D is false because trained algorithms need to be tested in an entirely independent dataset.
Reference: Computational pathology fundamentals
Comment Here