ThalNet: Deep Learning for Thalassemia via Blood Image Structure Function Image

Thalassemia, a genetic blood disorder, stems from faulty synthesis of a- or ß-globin subunits of hemoglobin, resulting in chronic hemolytic anemia. Diagnosis via complete blood cell (CBC) can identify Thalassemia but does not pinpoint specific anemias which require further tests that are often inaccessible in smaller health care settings due to equipment costs. We propose ThalNet, a more accurate and automated prediction model to screen individuals with Thalassemia from healthy individuals and classification between the two
$\beta$
- Thalassemia subtypes. We compared ThaINet against classical machine learning models and a recent deep learning model. In screening individuals with Thalassemia from healthy individuals using the ThaINet model, we found that the ThalNet model achieved an accuracy of 0.86, 0.88 precision, 0.87 recall (sensitivity), 0.78 specificity, and 0.85 fl-score. For the subtype classification of 13-Thalassemia, the ThalNet model achieved 0.67 accuracy, 0.71 precision, 0.67 recall (sensitivity), 0.67 specificity, and 0.64 fl-score. Our ThalNet model demonstrated superior performance over the ViT model in both Thalassemia screening task and the classification of
$\beta$
- Thalassemia subtypes task. Moreover, ThalNet outperformed the Classical approaches that utilize CBC count parameters, demon-strated its potential for clinical application in
$\beta$
- Thalassemia subtype classification using only a video of human whole blood sample.