Machine learning and machine teaching in histopathology.

Journal: Pathology, research and practice
Published Date:

Abstract

An artificial intelligence (AI) platform was trained by a consultant histopathologist to classify whole slide images (WSIs) of large bowel biopsies. Six medical students viewed WSIs of five large bowel biopsy cases and assigned the WSIs to one of the nine diagnostic categories. Then the students compared their answers with those generated by the AI. This training was repeated for a total of six rounds of five cases, and the accuracy of the students was recorded for each round. Each case had one or more WSIs. The student with the best final accuracy was asked to describe the morphological features that they had deduced. All the students improved during their training, from a mean accuracy of 13.7% in the first round to a mean accuracy of 77.1% in the sixth round (p = 0.0011). The student-deduced diagnostic features were mainly accurate. Some students learned more quickly than others.

Authors

  • Amy Louise Stokes
    University of Exeter Medical School, St Luke's Campus, Heavitree Road, Exeter EX1 2LU, UK.
  • Frederick George Mayall
    Department of Cellular Pathology, Musgrove Park Hospital, Taunton TA1 5DA, UK. Electronic address: Fred.Mayall@SomersetFT.nhs.uk.