The aim of this data challenge is through modeling of publicly available image data for DR to address the following learning objectives:|
Students should be able to develop a CDS based on DL that analyzes image data such that they can:
- learn to understand a data science application domain (diagnosis of DR)
- prepare an image dataset for the modelling task
- build and train DCNN model to classification images
- evaluate, critically reflect on, and improve performance of DCNN
- integrate the DCNN model in a CDS system considering the advances and limitations of the technology and all ethical implications of the ultimate goal of the system.
- develop a presentation pitch for a potential owner/user of such a system that conveys the value that the system will deliver
The objective of the data challenges courses is to teach students how to set up, execute, and evaluate scientifically sound data analyses that are adequate for stakeholders (users, enterprise, and society) using existing and available data. |
The topic of Data Challenge 1 is to develop a data driven solution in the healthcare sector that, among others, has particular economical and societal impact.
In this course, the problem is well-defined in terms of available data and target requirements. There are also a well-defined set of stakeholders. Students are tasked to identify and execute the modelling and a suitable analysis approach by themselves, and then reflect on the adequacy of the analysis for different stakeholders. Furthermore, due to the nature of the application, the analysis incorporates important ethical considerations.
The solution involves support for a decision on diagnosis of diabetic retinopathy (DR). DR is a disease that affects blood vessels in the light-sensitive tissue called the retina that lines the back of the eye. It is the most common cause of vision loss among people with diabetes. Progression to vision impairment can be treated if detected in time. Currently, detecting DR is a time-consuming and manual process that requires a trained clinician to examine and evaluate digital color photographs of the retina.
The challenge in this course is to develop a clinical decision support system (CDS) that can (partially) automate the diagnosis of the DR using retina images.
There are significant challenges analyzing medical images presents significant challenges of processing high dimensional image data to discover weak signals that are correlated with the outcome. Due to the dimensionality and the highly non-linear relationship between the features many machine learning methods struggle to produce models with acceptable performance. Recent developments in Deep Convolutional Neural Networks (DCNN) have been very successful in addressing these challenges.
However, beyond the technical challenges the task also requires good understanding of the domain, application of data science methods in clinical settings, particularly the ethical implications of using machine learning models in healthcare.
Furthermore, DL methods typically deliver ‘black box’ models that are difficult to interpret and are optimized to perform on a given training and test data, such that their performance in deployment may be not be fully predicable.
For further information see the course description at mytue.tue.nl.