Automatic analysis of thorax-abdomen CT scans

Patients with cancer typically receive CT scans before and during treatment to monitor the progression of their disease. These scans usually cover the complete body, except for arms, legs and head. In our hospital we make about 2000 of such scans every year. Radiologists need to carefully go over these scans, and compare them with the prior scans of the same patient. They need to look for any lesion, and measure the lesions to see if they grow or shrink. This is a very time-consuming task.

On top of that, there is a growing number of rules and evaluation criteria (RECIST 1.1, irRECIST 1.1, mRECIST) which complicates the reporting of these scans. RECIST (Response Evaluation Criteria In Solid Tumors) 1.1 is the most important guideline and describes how to assess the change in tumor burden. It includes the measurement of five target lesions (two per organ) and describes how to measure these. On follow-up scans, the same target lesions need to be measured again. Based on the measurements, the tumor response can be categorized as complete response (CR), partial response (PR), progressive disease (PD), or stable disease (SD). This is also a time-consuming task.

In a large project at DIAG we are developing an optimized software environment that assists radiologists in reading these scan more accurately and more quickly. This is a big project that involves many sub-tasks for which we continuously welcome support from MSc and BSc students.

A large database of thousands of annotated scans is available. This allows us to train complex deep learning systems to detect anatomical structures, lesions, and changes over time.


The work will be executed in the Diagnostic Image Analysis Group (DIAG), part of the Radiology and Nuclear Medicine department of Radboud University Medical Center Nijmegen.


At the moment we are looking for student who will segment all bony structures in thorax-abdomen scans. The adult human body has 206 bones. Of these, we find 60 in thorax-abdomen scans: the cervical vertebrae (7, are usually not all visible), thoracic vertebrae (12), lumbar vertebrae (5), sacrum and coccyx (2, but we treat them as one structure), sternum (1), ribs (24, in 12 pairs), humeri (2), scapulae (2), clavicles (2), hip bones (2), femural bones (2). We have a data set available in which these bones have all been manually segmented, and we have obtained promising first results using a 3D U-net. There are many challenges left though, such as correctly labeling all identical structures (the vertebrae, the ribs), dealing with abnormalities, missing structures and anatomical variants, and improving the 3D U-net approach.


  • Students with a major in computer science, biomedical engineering, artificial intelligence, physics, or a related area in the final stage of master level studies are invited to apply
  • Affinity with programming is required, interest and experience with machine learning preferred