Automated Detection of Blast Cells in Hematological Malignancies using Inception-based Residual Network

Detection and classification of malignant leukocytes is an important step for diagnosing acute myeloid leukemia (AML) using peripheral blood smear imagery. A survey of literature on AML shows that deep learning models are introduced for detection and classification of malignant leukocytes from peripheral blood smear imagery. However, deep learning-based classification models are sensitive to handle class imbalance in classification of malignant leukocytes. This paper proposes a cost-sensitive learning with fine-tuned transfer learning methodology for detection and classification of AML. During finetuning, a transfer learning model with cost-weights are introduced for each classes of malignant leukocytes to handle the bias and inaccurate predictions. Optimal transfer learning methodology and its parameters are selected based on the hyperparameter tuning approach. Model performances are shown using a dataset of more than 18,000 leukocytes. Fine-tuned cost-sensitive InceptionResNetV2 transfer learning model showed a higher performance compared to the other models with an enhanced precision 1%, recall 3%, and f1-score 3%. The proposed model can be used by a healthcare professional to enhance the early diagnosis of AML.