Comparative Analysis of Deep Learning Architectures for Blood Cancer Classification

An increased growth in the blood cancer necessitates the development of efficient, cost effective, timely, and accurate diagnosis. Traditional diagnosis methods are often invasive, expensive and time-consuming. The rapid artificial intelligence (AI) assistive advancement in the digital healthcare permits the realization of effective solutions in this framework. Specifically, the deep learning (DL), seems promising in an automated diagnosis. However, still a critical gap needs to be covered by understanding that which DL architecture performs better for the blood cancer detection. To address this crucial need, this paper presents a comprehensive comparative analysis of the key DL methods, used in an automated categorization of the blood cancer. The considered DL architectures are the MobileNetV2, DenseNet121, VGG16, ResNet50, and InceptionV3. The applicability is tested using two blood cancer datasets namely the Acute Lymphoblastic Leukemia (ALL) dataset and American Society of Hematology (ASH) dataset. Each model is meticulously trained and evaluated on the ALL dataset for binary classification and the ASH image bank for multi-class classification. The categorization performance is evaluated based on accuracy, precision, recall, F1 score, and latency. Results have shown an out performance of the MobileNetV2 compared to the counter DL architectures with a mean accuracy of 91.26%, precision of 92.94%, recall of 91.27%, F1 score of 90.58% and latency of 104.16 mins for ALL dataset and 88.11 accuracy, 90.23% precision, recall of 88.11%, 87.98% F1 score and latency of 11.16 mins for ASH dataset.