Student for improving DL model failure prediction in radiotherapy

Are you interested in optimising the implementation of AI in a real-world clinical workflow? Do you want to contribute to research that assists in developing automated quality-assurance tools for DL models in radiotherapy?

This opportunity is for a student role focused on improving DL model failure prediction in radiotherapy.

What will you do?

To optimise the implementation of DL models in the clinical radiotherapy workflow, we are developing an automated quality assurance (auto-QA) system that can automatically detect when the model makes a mistake and what the consequences of these mistakes are for the patient’s treatment plan. One of the elements of this system will be to detect when a DL segmentation model is applied to data outside its training distribution.

• You will conduct an overview of recent literature on promising out-of-distribution detection methods
• You will implement, analyse, and compare different methodologies to detect out-of-distribution inputs to a DL segmentation model (e.g. generative AI, uncertainty quantification, model feature distance, etc.)
• You will collaborate with the research team on how to integrate this element in the greater auto-QA tool
• You will contribute to the usability of AI by improving the implementation of AI in a clinical workflow
• There is room for your own input
• There is interest to turn the results into an academic publication

What is this research about?

At the department of radiotherapy, a radiation treatment plan is made for every individual patient based on imaging scans. The tumour and organs at risk are delineated to optimise dose to the target and minimise dose to healthy surrounding tissue. However, manual delineation is time-consuming, so we have implemented automated delineation with a deep learning (DL) model since 2018. But DL segmentation models do not have perfect accuracy for all patients and structures. Therefore, all DL segmentations have to be evaluated by clinicians. This partially diminishes the time-efficiency of the DL model implementation. Furthermore, this is a limiting factor in the adoption of online adaptive treatment, where a new treatment plan is made for each patient based on daily imaging.

Therefore, there is a growing interest in methods to automatically assess the quality of DL segmentation and provide this confidence indication to the clinicians that evaluate the segmentations. Recent research within the UMCG and other hospitals focuses on developing an auto-QA system. This system will consist of multiple layers that detect when and where the model makes errors and what the consequences of these errors are for the patient’s treatment plan.

At the moment, we have developed a reliable method for detecting local mistakes of a model. However, we are also interested in adding an extra layer to the auto-QA system that detects if the model is applied to the right patient, meaning whether the patient was within the training distribution of the model. Recent literature suggests multiple methods that could be suitable for out-of-distribution detection.

Therefore, we are looking for a master thesis student who wants to contribute to this research.

What do we ask?

For this project, we are looking for a university master’s or bachelor’s student with, for example, a background in artificial intelligence, computing science, applied mathematics, or a similar field.

• You have an interest in clinically applicable research
• You have experience with ML/DL modelling
• You are able to independently conduct research and have strong writing skills
• You enjoy collaborating with team members to develop innovative solutions

What do we offer?

• Internship agreement with UMCG
• Good supervision at UMCG
• Scientific working environment (AI in radiotherapy group)
• In consultation, you can partly work from home.