Publication

Abstract : Disease diagnosis classifies patients’ health conditions into specific grades and helps to make appropriate decisions for better treatment. Advancement in the field of microscopy and computer vision enables efficient disease diagnosis, a requirement for economical healthcare. Recently, deep learning recasts the face of computer vision, which outperforms humans in object recognition tasks. Usually, convolutional neural networks (CNNs) are used for image recognition tasks due to the fact that such a network architecture considers the spatial structure of the images. Instead of depending on CNN alone, here we introduce a new architecture, which consists of a shallow CNN appended with a single recurrent layer. Performance comparison of the proposed architectures on microscopic images have been done by using three different types of recurrent layers, such as recurrent neural network, long short term memory, and gated recurrent unit. We also evaluated the performance of all these models on three different disease diagnosis tasks from microscopic images: tuberculosis in sputum samples, intestinal parasite eggs in stool samples, and malaria in thick blood smears. In all cases, the proposed models produce better performance than state-of-the-art models. The proposed deep architectures for disease diagnosis fewer trainable parameters when compared to the existing state-of-the-art deep architecture.