An efficient deep learning system for automatic detection of Acute Lymphoblastic Leukemia

Early and highly accurate detection of rapidly damaging deadly disease like Acute Lymphoblastic Leukemia (ALL) is essential for providing appropriate treatment to save valuable lives. Recent development in deep learning, particularly transfer learning, is gaining a preferred trend of research in medical image processing because of their admirable performance, even with small datasets. It inspires us to develop a novel deep learning-based leukemia detection system in which an efficient and lightweight MobileNetV2 is used in conjunction with ShuffleNet to boost discrimination ability and enhance the receptive field via convolution layer succession. More importantly, the suggested weight factor and an optimal threshold value (which is experimentally selected) is responsible for maintaining a healthy balance between computational efficiency and classification performance. Hence, the benefits of inverted residual bottleneck structure, depthwise separable convolution, tunable hyperparameters, pointwise group convolution, and channel shuffling are integrated to improve the feature discrimination ability and make the proposed system faster and more accurate. The experimental results convey that the proposed framework outperforms others with the best detection performances. It achieves superior performance to its competitors with the best accuracy (99.07%), precision(98.00%), sensitivity (100%), specificity (98.31%), and F1 score (0.9899) in ALLIDB1 dataset. Similarly, it outperforms others with 98.46% accuracy, 98.46% precision, 98.46% specificity, 98.46% sensitivity, and 0.9846 F1 Score in ALLIDB2 dataset.