Robust identification of the L3 vertebra

Start date: February 1, 2023
End date: July 31, 2023

Clinical problem

Body composition is an important biomarker in the treatment of cancer. In particular, low muscle mass has been associated with higher chemotherapy toxicity, shorter time to tumor progression, poorer surgical outcomes, impaired functional status, and shorter survival. However, because CT-based body composition assessment requires outlining the different tissues in the image, which is time-consuming, its practical value is currently limited. The different tissue types are often outlined manually and only in a single slice of a CT scan at the level of the third vertebra (L3). This is based on the observed linear relation between cross-sectional areas and total body volume of muscle and fat. However, L3 localization can be challenging because of many frequently occurring anatomical variants of the lumbar spine. More onformation can be found on Radiopaedia here and here.

For use in both routine care and in research studies, automatic localization of the correct lumbar vertebra is desirable, especially if that would enable reliable and reproducible quantification of various tissue types on CT scans for body composition assessment.

For this clinical problem, and for others as well, accurate labeling of the vertebrae onmn CT scans is of great importance. In our group, a well performing method to segment vertebrae has been developed. It is available on grand-challenge. This method, howver, is not always accurate in labeling the vertebrae, especially not when anatomical variants are present. The goal of this project is to develop a new labeling method, given a CT scans on which one or more vertebrae are visible and a segmentation of all vertebrae is provided. The central slice of each vertebra should be provided and it is especially important that the middle slice of L3 is correctly determined.

Solution

There are various approaches one can take to solve this problem. One option is to label a single vertebra with high accuracy, and label the vertebra above and below it using their normal order, and detecting the presence of anatomical variants and updating the labeling based on that. You should investigate several options to tackle the problems and implement some of the most promising approaches.

Data

We have a large collection of over 100k CT studies available containing vertebrae. The radiology reports are also available. From this data set, cases with anatomical variants can be detected using natural language processing.

Results

An important aspect of this project is to provide the implementation as an easy to use web-app available here so that the software can be used by clinical researchers. It may also be interesting to set up a challenge on grand-challenge.org and/or write a publication.

Embedding

You will be embedded in the Department Of Medical Imaging at Radboudumc. We provide access to Sol, our large GPU cluster, and the cloud-based compute infrastructure of grand-challenge.org, powered by Amazon Web Services.

Requirements

  • Students Artificial Intelligence, Data Science, Computer Science, Bioinformatics, Biomedical Sciences or similar in the final stages of their Master education.
  • You should be proficient in Python programming and have a theoretical understanding of deep learning
  • Basic biological / biomedical knowledge is preferred.

Information

  • Project duration: 6 months
  • Location: Radboud University Medical Center

People

Oscar Esteban

Oscar Esteban

Ajay Patel

Ajay Patel

Coordinator RTC Deep Learning

RTC Deep Learning

Alina Vrieling

Alina Vrieling

Assistant professor

Health Evidence, Radboudumc

Matthieu Rutten

Matthieu Rutten

Associate Professor, Radiologist