Fig. 3: A high diagnostic accuracy of our AI models with the use of smaller datasets.

a Hematoxylin and eosin (H&E)-stained formalin-fixed paraffin-embedded tissue sections prepared from lymph nodes of DLBCL and non-DLBCL patients from four unrelated hospitals (hospitals A, B, C) were photographed (hospitals A, C) or scanned (hospital B) at ×400 original magnification to produce pathologic images for generating three separate AI models (Models A, B, C), each of which was specifically generated using the DLBCL and non-DLBCL samples from the corresponding hospital. A high diagnostic accuracy was reached by the three AI models (100% for hospital A, 99.71% for hospital B, 100% for hospital C, respectively). b Analysis of whole-slide images from hospital B from each patient by randomly selecting nine pathologic images within the DLBCL cell-containing areas. Thus, each experiment was done nine times. c The diagnostic accuracy dropped from 100 to 90.50% or 82.09% with or without unifying the shape of the images between hospital A and hospital C when cross-hospital use of the deep-learning model A was carried out to read the slide images of patients from hospital C. The diagnostic accuracy increased to 100% when the model A was used to read new images of patients in the same hospital. 100% diagnostic accuracy was also achieved when the model B was used to reach the slide images of patients from a new hospital (hospital D) after elimination of the technical variability introduced by slide preparation procedures and image collection equipment.