A new artificial intelligence (AI)-powered diagnostic tool, DeepHeme, is able to perform individual cell classifications with an accuracy comparable to human experts, according to a recent study published in Science Translational Medicine.
DeepHeme, a snapshot ensemble deep-learning classifier based on the ResNext-50 convolutional neural network architecture, was trained to classify 23 morphological types of hematopoietic cells from patient bone marrow aspirates (BMAs).
The training data consisted of 30,394 images from 40 patients derived from 400×-equivalent whole-slide images from the University of California, San Francisco (UCSF).
For validation, DeepHeme was tested on an independent dataset from a different institution, the Memorial Sloan Kettering Cancer Center (MSKCC). The validation dataset included 2,694 images from 10 morphologically normal patients and 11,076 images from 655 patients with normal or diseased marrow, scanned using a different whole-slide image system from the UCSF data, demonstrating the robust generalizability of the classifier.
The researchers compared DeepHeme’s diagnostic performance with that of three medical experts from different academic hospitals, demonstrating that DeepHeme achieved an accuracy comparable to, or exceeding, that of human experts, according to the investigators.
The senior author of the paper, Gregory Goldgof, MD, PhD, MS, is a computer scientist and pathologist and Director of AI and Computational Hematopathology at MSKCC.
“DeepHeme can analyze both blood and bone marrow samples and may support future efforts to improve personalized medicine,” Dr. Goldgof said in a press release from MSKCC. “Automation with DeepHeme will support large-scale research efforts to develop tools that better predict which patients will respond best to different treatment options.”
According to the press release, MSK plans to begin using DeepHeme in the clinic after further validation and may license it to other hospitals.
A demo of DeepHeme is available at https://www.hemepath.ai/.
