Automated Classification of Cells from Bone Marrow Cytology with Deep Learning

The manual classification of bone marrow (BM) cell morphology, a pivotal aspect of haematological diagnosis, is performed thousands of times daily due to the lack of comprehensive data sets and trained models. This process is often time-consuming and susceptible to human errors. The effectiveness of deep learning algorithms in biomedical applications is proven. The impact is undeniable and unquestionable as these techniques use extensive datasets encompassing diverse disease classes, meticulously annotated by medical professionals. This also eliminates any scope for error, shortens the diagnosis time and enhances the accuracy. Over time, deep learning techniques have improved further with new advancements. This research aims to evaluate the performance of several pretrained Convolutional Neural Network models to automate the classification of BM cells. More specifically, the models are compared by their ability to assist medical professionals in identifying the presence of ‘hairy cells’ in the bone marrow smear of a subject. The presence of hairy cells in the blood is indicative of the possibility of a person having Hairy Cell Leukemia. A custom CNN architecture, ConvNeXtSmall, ConvNeXtTiny, DenseNet121, EfficientNetB5, ResNet50, VGG16 and Xception were compared. ConvNeXtSmall has the highest validation accuracy of 0.85. DenseNet121 and ConvNeXtSmall have the highest F1 score of 0.81 while classifying Hairy Cells.