Deep Learning

RSIP Vision is one of the companies behind the wide adoption of deep learning techniques in the image processing and computer vision projects in the industry. Several reasons explain deep learning's popularity, the most important of which is obviously its capability to achieve better results than previous state-of-the-art techniques. Deep learning can be applied to many scientific areas and its exceptional performances (when correctly used) are often crucial to progress: think, for instance, how precious even a slight improvement can be in the medical or in the automotive fields, in terms of better health and increased security. Here are a few examples of successful projects by RSIP Vision, all using deep learning.

Deep Learning and Convolutional Neural Networks

RSIP Vision is widely using deep learning and Convolutional Neural Networks (CNN) in classification projects. We have decided to write this article to help data scientists and machine learning practitioners become more acquainted with the concept of deep learning and learn what the buzz is all about. We provide reference and links to authors and researchers who have explored the theoretical side of this powerful method. On our side, we apply it day by day to solve all kind of challenges encountered by our clients.  Read more...

Chromosomes Classification with Deep Learning

Chromosomes contain most of the DNA of living organisms. Karyotype tests are done to identify and evaluate size, shape, and number of chromosomes in the body cells. Cytologists able to perform manual chromosome classifications are rare and their task is performed much faster by machine learning systems. RSIP Vision developped an automatic chromosome classification algorithm using convolutional neural networks (CNN), a deep learning technique which enabled us to give our client a very robust chromosome classification which will keep improving with additional learning. Read more...

Glaucoma Detection by Deep Learning

Glaucoma, a high intraocular pressure (IOP) pathology, leading to damage of the optic nerve, can be better detected using deep learning techniques. When it detects the optical disc (the visible section of the optic nerve), the deep learning algorithm helps assess glaucoma in an automated way, starting from the region of interest and providing a reliable probability for the disease, which the physician will use to support both diagnosis and treatment decisions. Read more...

Deep Learning for Optical Character Recognition

Optical Character Recognition (OCR), used for the visual inspection of documents has found wide application in both industry and research. Convolution Neural Networks (CNN) have been found very powerful to detect text and recognize characters: like in the human visual system, different neurons and processing layers are sensitive to different features of objects: edges of objects, color gradients and other features. Deep learning is very precious in explaining the high diversity of OCR challenges: different languages, different handwriting, uneven illumination and so on. Read more...

Pulmonary Lymph Nodes Detection using CNN

Pulmonary lymph nodes provide precious information for lung cancer diagnosis, as well as for staging and evaluation of treatment efficacy. Thus, automatic and accurate segmentation and detection of enlarged lymph nodes in the lungs is a task which offers great clinical value. To achieve the highest accuracy, it is recommended to train a deep learning classifier. Convolutional Neural Network (CNN) are used to classify the candidate voxels with dramatic improvement over previous state-of-the-art results. Read more...

Deep Neural Networks for Vessel Segmentation - Fundus

Various segmentation methods, whether based on Convolution Neural Networks or traditional image processing techniques, can be used to delineate the vascular tree in clinical imaging. Given the few features distinguishing veins from arteries (usually brighter and thinner than veins), the challenge consists of training a binary classifier assigning each pixel to the category of vein or artery. This article covers the advantages of using CNNs and deep neural networks for the classification and segmentation of vessels in fundus images. Read more...

Automated Lung Nodule Classification with Deep Learning

Lung cancer early detection is a vital task which is made difficult by the small size of pulmonary nodules, the detection of which on thousands of CT scans every day is excessively time-consuming. Computer-aided lung nodule classification can dramatically boost the speed of diagnosis. Recommended solution starts from bidimensional images obtained from CT scan and displaying suspicious nodules areas: these are inserted into an autoencoder, from which two hundred dimensional features are extracted. These learned features are then confronted with a trained classifier to produce the final lung nodules classification. Read More...

Cyst Detection Using Convolutional Neural Networks

The goal is to automatically detect the appearance of Cystoid Macular Edema (CME) in Optical Coherence Tomography (OCT) images. The deep learning technique used, Convolutional Neural Networks, takes as an input patches of pixels from within the retina. These patches were generated from previous segmentation of retinal images. A further segmentation of the retina is performed using an image processing algorithm called SLIC. Every superpixel thus generated, after being labeled as in the OCT scan, is fed into the neural network to detect the cyst. Read more...