Comparative Analysis of Transfer Learning Strategies for Automated Identification of Bone Marrow Cell Morphologies

The gradual decrease in the average lifespan of mankind is considered to be caused by various diseases, along with hematological disorders are prominent. Diagnosis of this hematological disorder requires an understanding of the morphology of bone marrow cells. This study introduces a computer vision-aided solution in order to perform an automated evaluation of bone marrow smears from the examiner’s skill and experience. In this regard, we employ the publicly available dataset of 170,000 expert-annotated cell images from the bone marrow smears with twenty-one separate groups. A comparative study of the models developed from our dataset has been performed using eleven customized deep-transfer learning algorithms. The EfficientNetB3 outperforms other existing models. We have achieved the maximum accuracy of 87.2% by EfficientNetB3 compared with other deep learning models. Similarly, the recall in EfficientNetB3 was 87.22%, the precision was 87.11%, and the F1 score was 87.15%. Conversely, DenseNet201 offers the second-most effectiveness, with an accuracy of 86.32%.