Project DeepScreenAMD

Our Goals

The main goal of the project is to design, assess, and extensively interpret a deep learning (DL)-based decision support system aimed at early, accurate detection of AMD. We will focus both on detecting the presence of the disease and on grading its different stages. To achieve this goal, different architectures from diverse DL approaches will be implemented. Furthermore, in order to interpret, from a clinical point of view, all the detected risk factors leading to the detection of AMD from fundus images, a thorough analysis of every model will be conducted using XAI techniques.

Specific objectives

Objective 1

To build up a private database composed of novel fundus images. This goal includes the capture, organization, and manual annotation of the images.

Objective 2

To design and optimize high-performance models for AMD diagnosis based on novel individual DL approaches.

Objective 3

To explore and develop novel architectures by combining different DL-based approaches.

Objective 4

To evaluate the detected patterns inherent to input fundus images linked to the presence and severity of AMD and its consequences.

Objective 5

To interpret and explain the diagnostic decisions of the DL models based on the detected risk factors.

Expected impact

Scientific-technical impact

Development of novel high-performance decision support-systems based on the emerging DL techniques.
Identification of novel image landmarks in fundus images for an early detection and grading of AMD.
Clinical analysis of the results through the extensive interpretation of automated models.
New software applications for ophthalmology research with potential to be integrated in the devices of leading manufacturers.

Social and clinical impact

Reducing waiting lists in ophthalmology.
Development of automatic methods that contribute to the early diagnosis of AMD and that help prevent the evolution of dry AMD to wet AMD.
Gaining insight into the pathophysiology of AMD through the extensive interpretation of automated models.
Evaluation of the final diagnostic/management methodologies in clinical practice with real patients.
Publishing the results in a website, social media, open access repositories and scientific divulgation magazines to maximize the dissemination in the general population.
Employment creation.

Economic impact

Decreasing the total number of appointments with the specialist and associated direct costs.
Savings in human and material resources, allowing to be assigned to more demanding tasks.
Savings in indirect costs related to health, since early diagnosis and adequate therapy would improve the quality of life of people with AMD.

Knowledge transfer and results exploitation

Action 1

Identification of results with high potential to be exploited/transferred, including but not limited to: (i) New algorithms (software) to detect AMD based on DL methods; (ii) New algorithms (software) for the management of patients with AMD based on XAI methods.

Action 2

Involving possible observing promoters in the publication and dissemination tasks.

Action 3

Assessment of the feasibility of transferring the automated decision-support systems to (i) clinical practice environments and (ii) companies in the field of ophthalmological services/therapies, in order to provide added value to the patients.

Action 4

Collaboration with the Office for the Transfer of Research Results (OTRI) from the UVa, which will provide advice and support on different ways of transferring and exploiting research outputs.

Action 5

Identification of the scientific-technical knowledge with potential for being protected by means of intellectual property rights. Assessment of the feasibility of applying for software licenses/patents.

Action 6

Development of a careful market analysis in order to identify strengths and opportunities, as well as potential barriers and threats to exploit or transfer the results. This would include the contact with health institutions, private companies, and service providers in the field of ophthalmological medicine.

Project tasks

Design and creation of the project database

Analysis of the public AREDS database. Creation of the proprietary database of retinal images from a clinical setting. Manual annotation of the images of the database, including the clinical signs of AMD and AMD severity in each image. Design of validation strategies.

Bibliographic review of automatic methods

Compilation and analysis of the latest scientific studies related to DL and XAI. Comparison between methods.

Implementation of selected automatic methods

Implementation, using Matlab® and Python tools, of the selected DL architectures and XAI procedures.

Automatic detection and grading of AMD

DL training and validation for all images of the database for the detection and grading of AMD. Study of the image features that potentially differentiate AMD landmarks using XAI.

Statistical analysis of the results, discussion and drawing of conclusions

Study and application of the most appropriate statistical methods to detect AMD and grade its severity in patients. Study of the most relevant image landmarks in the context of AMD detection and grading. Interpretation of the results from a clinical point of view. Extraction of the main conclusions of the study.

Dissemination and transfer of results

Dissemination of preliminary results in conferences. Publication of papers on the results of the project in prestigious international journals. Preparation of dissemination material: website, social networks, organization of seminars, awards, and press releases. Preparation of reports on the evolution of the project for the Observing Promoter entities (OPs): University of Porto, Hospital Clínico Universitario de Valladolid. Annual meetings to discuss the results. Preparation of reports on the development of patents.

Management and coordination

Coordination and control of each of the tasks and subtasks.