Evaluation of an expert system on lymph node pathology
Related Experiment Video
Updated: Oct 31, 2026
Comparison of Predictive Performance of Three Lymph Node Staging Systems in Colorectal Signet Ring Cell Carcinoma Based on Machine Learning Model
Published on: April 18, 2025
B N Nathwani1, K Clarke, T Lincoln
1Department of Pathology, University of Southern California School of Medicine, USA.
Background:
Pathfinder is an Expert System that assists pathologists in making accurate diagnoses in the domain of lymph-node pathology. Pathfinder provides a differential diagnosis based on the initial histological feature(s) observed by the pathologist, and suggests to the user additional histological features for observation that are likely to narrow the differential diagnosis.
Purpose:
To evaluate the diagnostic accuracy of pathologists with and without the Pathfinder.
Methods:
Thirty H&E stained slides from 30 lymph node biopsy specimens on which a Consensus diagnosis was made by two experts were reviewed by 19 pathologists to evaluate Pathfinder. After a period of training, 10 pathologists using Pathfinder (Interactive Computer Method) and 9 pathologists using the Routine Method (diagnosis without computer) determined a differential diagnosis for 15 slides (Test 1). Pathologists were then crossed over, trained, and evaluated the remaining 15 slides (Test 2). For each test, the proportion of "correct" diagnoses was compared between methods. In addition, the information integration attributes (making logical diagnosis given a set of specific histologic features) of Pathfinder and pathologists were compared. Finally, feature identification and quantification skills of pathologists were determined and correlated with the percent correct diagnosis.
Results:
The diagnostic accuracy using Pathfinder was greater than that using the Routine Method (40% v 32%, P = .02). Diagnostic accuracy for the group of pathologists who made diagnosis using the Routine Method in Test 1 increased when they made diagnosis using Pathfinder in Test 2 (27% to 44%, P < .0001). The proportion of correct diagnosis for the group of pathologists who used Pathfinder in Test 1 remained virtually unchanged when they gave up Pathfinder in Test 2 (35% to 37%). The percentage of incompatible feature identification ("atypical proliferation" diagnosis) was significantly lower after using Pathfinder (P < .0001). In addition, information integration attributes of Pathfinder were significantly superior than that of the pathologists (P < .0001).
Conclusions:
Pathfinder is a valuable tool that assists pathologists in making accurate diagnosis because it has superior attributes than pathologists to integrate information and to screen for observations incompatible with any specific disease.
