Advanced search
Start date

Preventive medicine by means of deep learning techniques applied in healthcare prognosis

Grant number: 18/17620-5
Support type:Scholarships abroad - Research
Effective date (Start): February 16, 2019
Effective date (End): December 15, 2019
Field of knowledge:Physical Sciences and Mathematics - Computer Science
Principal Investigator:José Fernando Rodrigues Júnior
Grantee:José Fernando Rodrigues Júnior
Host: Sihem Amer Yahia
Home Institution: Instituto de Ciências Matemáticas e de Computação (ICMC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
Local de pesquisa : Laboratoire d'Informatique de Grenoble (LIG), France  


Deep Learning (DL) describes a class of algorithms capable of combining raw inputs into successive layers of intermediate features to achieve computational intelligence. These algorithms have shown impressive results across several domains. In Medicine, for example, which is a data-rich discipline, the data are complex and often not yet understood. Deep Learning techniques may be particularly well-suited to solve problems in this field. In that context, the aim of this project is to explore the possibilities of DL in the context of computer-aided medicine; the activities include dealing with the various problems related to the application of DL to the specific context of clinical data. The issues include pre-processing demands to produce large, labeled, and cleaned datasets of clinical data; modeling complex information according to the demands of DL processing; fine-tuning DL architectures with respect to the specific problems of prognostic care; iterative training-testing rounds to achieve highly accurate methods; clinical validation of the results; and dissemination of the methods in the form of real-world applications with actual social impact. These tasks will be carried out over two initial problems: (i) the automatic detection of skin tumors from skin lesion photos; (ii) the prognostic of patients based on the clinical history as given by Electronic Medical Records (EMRs). The first problem will use open-access data from the ISIC Dermoscopic Archive, and from the Edinburgh Dermofit Library to transfer-learn over the ResNet-152 network modeled to the ILSVRC competition. The second problem will use data from the French institution AGIR to guide the modeling and training of a Long Short-Term Memory DL architecture capable of inferring over large contexts of sequential information, as those seen in EMRs. The two methods shall impact on practices of preventive medicine, allowing for early detection of skin tumors and early recommendation of treatments/procedures. This impact is of special importance to the Brazilian scenario, in which basic health care is inaccessible in many regions of the country. The postdoc period will contribute to the expertise of the proponent in a currently active research field, promoting new investigative fronts to his research group. (AU)