Application of Convolutional Neural Network Image Analysis and Machine Learning to Basic Blood Tests for Intelligent Diagnostic Assistance.

Background and objectives: We developed an automated morphological image recognition deep learning system (image recognition DLS) of peripheral blood cells, then constructed the diagnostic assist DLS combining image recognition DLS data with complete blood count (CBC) data. This study aimed to evaluate the clinical performance of the image recognition DLS and the diagnostic assist DLS in routine examinations.

Methods: The image recognition DLS was trained using datasets containing 1 476 727 images of white blood cells (WBCs), nucleated red blood cells (NRBCs), and large platelets to differentiate 14 blood cell types and to recognize 24 morphological characteristics. CBC data were obtained through the automated hematology analyzer (Sysmex XN-9000) and combined with the image recognition DLS data to construct the diagnostic assist DLS. The clinical performance of the image recognition DLS was evaluated using 128 716 blood cell images from 589 smears obtained from healthy subjects, ALL, AML, ML, MPN, and MDS cases in routine examinations.

Results: The image recognition DLS classified 14 blood cell types with an accuracy of 97.3%-99.9%. The accuracy of 11 morphological characteristics exceeded 90%. Blast cells were detected accurately on all slides, where they were identified by manual microscopy. Malignant lymphocytes were classified as blasts and/or lymphocytes with the morphological characteristics of each subtype of lymphoma. The diagnostic assist DLS successfully differentiated MDS, achieving an AUC (area under the curve) of 0.99.

Conclusion: This study demonstrated the potential of the diagnostic assist DLS, utilizing morphological image recognition DLS data combined with CBC parameters, as a promising diagnostic tool.