Towards a Deep Learning Approach to Discriminate Hereditary Anemias

• Main Authors: Morabito, Francesco Carlo, Ieracitano, Cosimo, Mammone, Nadia, Valentino, Marika, Wang, Zhe, Schiavo, Michela, Bianco, Vittorio, Iscaro, Anthony, Nostroso, Antonella, Andolfo, Immacolata, Russo, Roberta, Miccio, Lisa
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Convolutional neural networks; Deep learning; Digital holography Microscope; Field programmable gate arrays; Hereditary anemias; Holography; Label-free Diagnostics; Microscopy; Phenotypes; Single cell imaging; Support vector machines; Technological innovation; Testing
• ISSN: 2687-6817
• DOI: 10.1109/RTSI61910.2024.10761680

Anemia affects more than 1.6 billion people globally; about 10% of these patients are affected by rare anemias, of which 80% are hereditary. Hereditary anemias (HA) are characterized by complex genotype/phenotype correlations and the current diagnostic workflow is made of three investigation lines: i) clinical tests (familial history, complete blood count, and peripheral blood smear); ii) specialized biochemical tests, and iii) genetic testing. Because of the diversity and overlapping phenotypes the clinical diagnosis could be not univocal and consequently prevent precise and correct care. The present study proposes a diagnostic step combining single-cell imaging and Artificial Intelligence (AI) to identify the macro-class of hereditary anemias. Indeed, the method is based, on one hand, on the development of a Digital Holographic (DH) system in microscopy configuration to image the Red Blood Cells (RBC) in static conditions. Such DH approach allows to retrieve the quantitative phase contrast maps of RBC that contain information on the morphology and on the hemoglobin content of the sample; on the other hand, it is based on the development of a custom Convolutional Neural Network (CNN) that exploits the quantitative phase information for classifying single RBC of healthy subject and of subjects affected by HA. Experimental results (i.e., average accuracy of 93.6%) encourage the development of an effective tool that would be integrated into the second line of the current diagnostic workflow in order to speed up diagnostic results and minimize error.