A combination of simple and dilated convolution with attention mechanism in a feature pyramid network to segment leukocytes from blood smear images

Leukemia, a type of blood cancer, is amongst the most deadly cancers worldwide. Since it affects leukocytes in the bloodstream, fast and early detection of abnormal leukocytes is required. Thus, precise detection of leukemia highly relies on accurate segmentation of leukocytes from blood smear images. The segmentation process has become quite robust with the development of deep neural networks, especially convolutional neural networks (CNNs). Such models have also shown superior results compared to traditional machine learning algorithms. This work represents a deep learning-based encoder–decoder model that focuses on salient multiscale leukocyte features. It is accomplished by combining features derived from standard and dilated convolutions. Using a convolutional block attention module (CBAM) in the network facilitates the extraction of refined features. We evaluated the performance of the proposed approach by conducting ablation studies on three publicly available datasets: ALL_IDB1, CellaVision and LISC. The first study is conducted to finalize the architecture of the dilated encoder path. In the subsequent study, a series of experiments are performed to obtain the most effective attention module. The last set of experiments deals with only a single encoder path that encapsulates dilated convolutions' importance. The resultant values of the proposed method are also compared with the state-of-the-art techniques using four performance indices: Dice score, IoU, PPV and NPV and qualitatively by visual results.