![deepfocus crop images deepfocus crop images](https://astrobites.org/wp-content/uploads/2014/07/lang_et_al-fig4-1024x599.png)
![deepfocus crop images deepfocus crop images](https://support.micasense.com/hc/article_attachments/115011238628/cropfocus.png)
This goal has been examined within the context of machine learning and CNNs in the past. Parasite, which is the most clinically relevant and deadly cause of malaria. We focus on the particular task of classifying if red blood cells are infected with the Plasmodium falciparum After presenting our "physical CNN" model, we then present simulations and experimental results for improved classification of cells from microscopic images. In this work, we first merge a general model of optical image formation into the first layers of a CNN. Alternative schemes like Fourier ptychography (FP, complex-valued reconstruction in b) can improve spatial resolution for better subsequent classification, but must capture more images and are thus less efficient. The classification accuracy with images from our optimal illumination technique in (c) notably exceeds that for a standard bright-field imaging setup in (a). Figure 1: We use a convolutional neural network to jointly optimize the physical layout of microscope illumination as well as a classifier to determine if cells are infected by the malaria parasite. We are hopeful that the presented framework will help bring together the growing field of deep learning with those who design the cameras, microscopes and other imaging systems that capture the training data that most learning networks currently use. By simply displaying two particular patterns on an LED array placed beneath our microscope, we can increase infection classification accuracy by 5-10%. We include a model of optical image formation into the first several layers of a neural network, which allows us to jointly optimize the design of a microscope and the various weights used for image classification in one step, thus forming a specific type of “classification microscope." As an experimental demonstration, we jointly determine an optimal lighting pattern for red blood cell imaging as well as a CNN classifier to test for infection by the malaria parasite ( Plasmodium falciparum). Here, we try to close the gap between how images are acquired and how they are post-processed by CNNs. Standard and alternative microscope lighting designs. Optimization technique with an experimental microscope configuration thatĪutomatically identifies malaria-infected cells with 5-10 The resulting network simultaneouslyĭetermines an ideal illumination arrangement to highlight important sampleįeatures during image acquisition, along with a set of convolutional weights toĬlassify the detected images post-capture. We increase the classificationĪccuracy of a microscope's recorded images by merging an optical model of imageįormation into the pipeline of a CNN. Physical layout of the imaging device itself. Neural network (CNN) not only to classify images, but also to optimize the To resolve with a standard optical microscope. Primarily transparent to visible light and contain features that are difficult However, many biological samples of interest are Deep learning algorithms offer a powerful means to automatically analyze theĬontent of medical images.