Acute Myeloid Leukemia subtype classification using single blood cell images.

Acute Myeloid Leukemia (AML) is a heterogeneous cancer that affects the myeloid lineage of blood cells. It requires accurate subtype identification for effective treatment. This study proposes a novel framework for AML detection and subtype classification using a two-step classification strategy: an initial classifier, trained on a dataset of single-cell peripheral blood smear images, categorizes individual cell images according to their cell type, followed by a model that utilizes these predictions to detect the presence of AML and determine the subtype. The methodology incorporates custom-trained CNNs and conventional machine learning algorithms to build an "ensemble of ensembles" which has been unified through a soft voting mechanism to enhance classification robustness. To address domain shift and variability in staining and contrast across datasets, cross-dataset validation and a robust preprocessing pipeline involving stain normalization and brightness correction have been tested and implemented. Additionally, Multiple Instance Learning and pseudo-labeling paradigms have been explored to compare their performance with the proposed methodology, which has demonstrated promising results and offers a practical foundation for further development in image-based AML subtype identification. The single cell classifier achieved 94.7% accuracy, 95.2% precision, and a 94.8% F1-score and our proposed ensemble model achieved 94% accuracy when distinguishing between healthy patients from AML patients.