ARTIFICIAL INTELLIGENCE, DEEP MACHINE LEARNING, AND IMAGE ANALYSIS IN DIAGNOSTIC HAEMATOLOGY: A SYSTEMATIC REVIEW

Abstract

Introduction: Laboratory haematology relies significantly upon the visual examination of routinely prepared publications as well as the histological or cytological investigation of bona-fide histological samples. As there exist a number of limitations to using human visual inspection to assess cytological or histological materials, there is an increasing interest in developing methods to assist with these tasks through the use of Artificial Intelligence (AI) technology, including Deep Learning (DL) systems. One such system is the Convolutional Neural Network (CNN) model. Objectives: To undertake a systematic review to critically evaluate the use of AI and DL techniques as they apply to the analyses of haematological samples, including: Classifying blood cells; Detecting malignant blood cells; Predicting prognosis for patients diagnosed with blood disorders; Determining methodologies to improve efficiency within the laboratory environment. Materials and methods: A systematic search in the literature was conducted, including all databases from inception to December 3rd of 2023, in accordance with the PRISMA 2020 guidelines. All original data-based publications that used AI or DL technology to aid in the diagnosis of haematological disorders, using either digital microscopy, flow cytometry, or imaging of bone marrow aspirates were included. The QUADAS-2 tool was used to evaluate the level of bias associated with the included study. Results: A total of 62 articles met the inclusion criteria. These studies included: 41 studies focused on making a visual assessment of bone marrow and/or blood cell morphology; 11 studies focused on making a visual assessment of bone marrow; and 10 studies focused on using a digital image of flow cytometry. The average accuracy of white blood cell (WBC) classification using a CNN-based model was found to be 95.5–99.4% (all five-part differential classifications), while the sensitivity for detecting acute leukaemia blasts ranged from 93.1 to 98.7% with a specificity of 91.1–99.1%. The DL model used for the diagnosis of sickle cell disease had a significantly higher AUC (0.996) than the ML model (0.921). The C-statistics found when using prognostic stratification models in myelodysplastic syndrome ranged from 0.78 to 0.84. Conclusions: AI and deep learning have been shown to have a high level of diagnostic accuracy in performing haematological morphological tasks, with the level of diagnostic accuracy approaching or exceeding that of human experts. In order to successfully develop clinically applicable AI-based technologies, there are several challenges such as clinical validation, regulatory approval, implementation, and equal access to AI-based diagnostics that must be addressed. There is an immediate need for interoperable and standardized reporting systems for AI diagnostic performance, as well as for the completion of additional multicentre prospective clinical trials before clinical implementation will occur.