Home > Informatics, digital & computational pathology > Computer aided detection

Informatics, digital & computational pathology

Artificial intelligence

Computer aided detection


Po-Hsuan Cameron Chen, Ph.D.
David F. Steiner, M.D., Ph.D.

Last author update: 14 December 2023
Last staff update: 14 December 2023

Copyright: 2019-2024, PathologyOutlines.com, Inc.

PubMed Search: Computer aided detection

Po-Hsuan Cameron Chen, Ph.D.
David F. Steiner, M.D., Ph.D.
Page views in 2023: 126
Page views in 2024 to date: 7
Table of Contents
Definition / general | Essential features | Key components | Terminology | Applications | Implementation | Advantages | Microscopic (histologic) images | Board review style question #1 | Board review style answer #1 | Board review style question #2 | Board review style answer #2
Cite this page: Chen PHC, Liu Y, Steiner DF. Computer aided detection. PathologyOutlines.com website. https://www.pathologyoutlines.com/topic/informaticscompaiddetect.html. Accessed November 27th, 2024.
Definition / general
  • System to assist with locating histologic or cytologic findings of interest
  • May be particularly relevant for
    • Small or sparse pathologic findings that could potentially be missed
    • Quantitation tasks in which miscounting may be likely
  • By contrast, a computer aided diagnosis system is meant to help with diagnostic interpretation, such as tumor grading or subtyping a histologic or cytologic feature (i.e., "what" as opposed to "where")
Essential features
  • System that assists with detection or localization of findings of interest
  • Requires an algorithm to detect findings and a user interface to assist the pathologist in the diagnosis
  • Typically is most useful for increasing sensitivity, efficiency or quantitation
Key components
Terminology
  • Computer aided detection (CADe): technology or system that detects findings of interest on medical images, typically with the goal of reducing the false negative rate
  • Computer aided diagnosis (CADx): technology or system that characterizes or interprets findings on medical images (such as tumor grade)
  • Model: the machine learning algorithm underlying the CADe system
  • Receiver operating characteristic (ROC) curve: a plot that represents the tradeoffs between sensitivity and specificity across a range of decision thresholds
  • Operating point: a specific point along a receiver operating characteristic curve, thus representing the specific diagnostic tradeoff of sensitivity and specificity at that point
    • Human diagnostic performance is often represented as a single operating point for each individual
    • Diagnostic algorithms often have continuous output (for example, a number between 0 and 1)
      • Thus, a threshold corresponding to a specific operating point must be chosen in order to decide which regions to highlight
Applications
  • Pap smear screening
  • Peripheral blood smear counting
    • One of the original CADe applications in pathology (Ann N Y Acad Sci 1966;128:1035)
    • Clinical task: differential cell count of peripheral blood smear slides
    • Algorithm methodology: convolutional neural networks, artificial neural networks (Clin Lab Haematol 2003;25:139, Am J Clin Pathol 2005;124:770)
    • Evaluation: comparison of classification accuracy between automated systems (CellaVision DiffMaster Octavia and DM96) and manual differential cell counts (J Clin Pathol 2007;60:72, Clin Lab Med 2002;22:299)
      • Both systems correlated well with manual counts
      • Shortened review time compared with an experienced clinical laboratory scientist
      • Both automated systems met the hematological lab requirements in terms of reliability and efficiency
  • Mitotic counting in breast cancer
    • Clinical task: mitotic counting for the histological review of breast cancer specimens
    • Algorithm methodology: decision trees that distinguish between mitosis and nonmitosis using image based statistical and morphological features (J Pathol Inform 2013;4:10)
    • Evaluation: automated versus manual and relabeling of manual reads (Med Image Anal 2015;20:237)
      • Error rate comparable to interobserver pathologist agreement
  • Classification of follicular lymphoma
    • Clinical task: centroblast counting for the histological grading of follicular lymphoma
    • Algorithm methodology: k-nearest neighbor classifier identifies high power fields of interest and classifies low versus high risk categories using a color coded map (BMC Med Inform Decis Mak 2015;15:115)
    • Evaluation: comparison of accuracy and interrater variability with and without model assistance
      • Improved accuracy, especially among residents
      • Improved consistency (reduced variability) across both experts and residents
      • Residents were more likely to change scores after using the computer system
  • Metastatic cancer detection in lymph nodes
    • Clinical task: finding metastatic breast cancer in sentinel lymph node biopsies
    • Algorithm methodology: convolutional neural network identifies and highlights tumor cells in lymph nodes (Arch Pathol Lab Med 2019;143:859, JAMA 2017;318:2199)
    • Evaluation: board certified pathologists digitally reviewed slides in assisted versus unassisted modes (Am J Surg Pathol 2018;42:1636)
      • Increased sensitivity for micrometastases
      • Shorter review time per image for micrometastases and negative images
      • Image review of micrometastases is considered to be easier with assistance
  • Prostate cancer detection in biopsy specimens
    • Clinical task: identifying prostate cancer in needle core biopsies
    • Algorithm methodology: convolutional neural network identifies tumor
    • Evaluation: pathologists review biopsies, initially without assistance and immediately after with algorithm assistance that provides binary classification of suspicious or benign and with location most likely to be cancer according to the algorithm (Arch Pathol Lab Med 2023;147:1178)
      • Increased sensitivity, small increase in specificity
      • Small increase in efficiency
  • PDL1 immunohistochemistry quantitation in lung cancer
    • Clinical task: quantifying PDL1 staining in lung cancer
    • Algorithm methodology: convolutional neural network to identify individual PDL1 positive cells on IHC image, followed by quantitation of positive cells
    • Evaluation: assisted and unassisted concordance of tumor proportion score (TPS) performed by 3 pathologists (Eur J Cancer 2022;170:17)
    • Additional evaluation against clinical outcome upon immune checkpoint inhibitor treatment and in additional cancer types (Mod Pathol 2022;35:1529)
      • Increased concordance with algorithm assistance and moderately improved prediction of response to treatment associated with the algorithm provided scoring subgroups
Implementation
  • Development: develop a computer algorithm to accurately identify a finding of interest, which typically involves
    • Collecting data: slides / images with expert labels for the findings of interest
    • Training model: supervised learning if expert labels provided
    • Evaluating model: using a set of images not used for training
  • Inference: application of the algorithm to the images of interest (i.e., running the algorithm on the image)
    • Requires digitized slides or computer integration with a microscope
    • Operating points (thresholds for algorithm output) are applied to determine which finding(s) should be highlighted (typically based on validation data)
  • Integration: integration with the clinical workflow is often a key challenge for an effective system and is likely unique for different uses
  • Evaluation: an accurate computer algorithm may not translate directly into a useful system; its effectiveness in improving clinical workflows or patient outcomes must be evaluated
  • Reference: Med Phys 2013;40:087001
Advantages
  • Improving efficiency: by highlighting regions of interest and reducing review time
  • Improving accuracy: by highlighting small or rare lesions or cells that are easily missed
  • Improving consistency: by helping to count histologic features of interest, such as immunohistochemistry quantitation, mitoses or centroblasts
  • Reference: Cancer Imaging 2005;5:17
Microscopic (histologic) images

Contributed by David F. Steiner, M.D., Ph.D.

Detection of metastatic tumor in lymph nodes

Board review style question #1
Detection of metastatic tumor in lymph nodes


Which of the following is an example of computer aided detection (CADe) in pathology?

  1. An algorithm that can accurately grade prostate cancer
  2. Comparison of algorithm performance to pathologist performance for classification of follicular lymphoma
  3. A system that uses an algorithm to identify mitoses on a digital pathology image and highlights those regions for review by a pathologist
  4. A tool that uses a machine learning algorithm to identify and report lymph node metastases without review by a pathologist
Board review style answer #1
C. A system that uses an algorithm to identify mitoses on a digital pathology image and highlights those regions for review by a pathologist. Computer aided detection (CADe) systems in pathology utilize an algorithm to identify features or regions of interest but also must communicate that information back to the human experts in a useful way. Answers A and B are incorrect because they both describe computer aided diagnosis (CADx), not detection. Answer D is incorrect because CADe can identify metastasis but still needs review by a pathologist.

Comment Here

Reference: Computer aided detection
Board review style question #2
Which of the following is a required element of a computer aided detection (CADe) in pathology?

  1. The ability to apply an algorithm to pathology images
  2. An algorithm that can identify all possible features of interest for a given image
  3. Choosing an operating point for the model that is superior to pathologists for both sensitivity and specificity
  4. A convolutional neural network (CNN) trained to classify pathology images
Board review style answer #2
A. The ability to apply an algorithm to pathology images. The only required element above is the ability to apply the algorithm to the image, either via slide digitization or computer integration with a microscope. Answer B is incorrect because computer aided detection (CADe) systems might only identify a single feature for a given image. Answer C is incorrect because other types of algorithms can be used and CADe systems do not necessarily need to be more sensitive or specific than pathologists in all cases in order to be useful. Answer D is incorrect because while CNN training can be useful, it is not necessary for all CADe systems.

Comment Here

Reference: Computer aided detection
Back to top
Home > Informatics, digital & computational pathology > Computer aided detection
Image 01 Image 02