Introduction
Liver metastases is a term coined to refer to tumors that spread to the liver from other primary tumors (i.e., colorectal cancer). According to the World Health Organization (WHO), colorectal cancer is a leading cause of cancer-associated deaths with liver metastases developing in 25–30% of those affected. In Norway, approximately 4000 new colorectal cancer cases are diagnosed every year. In general, approximately 11 000 controls are performed per year. Around 1400 patients develop liver metastases either during their treatment or follow-up period.
Various imaging techniques such as Computed Tomography (CT), Magnetic Resonance Imaging (MRI), and Positron Emission Tomography (PET) play an essential role in the diagnostic workup and post-treatment control, of a wide array of liver lesions. Humans mainly perform the interpretation of these medical images today. However, these images are challenging and time-consuming to analyze. A great number of these CT and MRI examinations are negative, in the sense that there is no pathology developing in the liver. Thus, a vast number of examinations are normal, but still just as time-consuming and challenging to interpret for radiologists as those showing pathology. Manual segmentation, detection, and interpretation are based on experience, which is subjective and can be prone to error.
AI is getting attention for medical imaging of the liver, and there are some progresses in this regard, however, with limited success in terms of transparency, external validation, and interpretability. However, current models still require further improvement in performance to be acceptable in real-life settings. This can be mainly attributed to the fact that pixel-level annotations are difficult, costly, and extremely time-consuming to obtain, which limits the use of completely supervised approaches to small datasets.
Research Agenda
The project aims to develop and validate AI and ML algorithms for effective and reliable automated examination of CT and MRI scans of livers with potential liver tumors. To effectively achieve the main objective, the project will provide solutions for the following subtopics within the framework of the main objective:
- Develop accurate deep learning systems for automatic detection of liver lesions in CT and MRI scans.
- Develop deep learning algorithms for the evaluation of liver lesions, i.e., systems that can accurately differentiate between different types of liver tumors.
- Validation of the deep learning systems in a complete regional rectal cancer cohort.
- Advance the current state-of-the-art within explainable AI and weakly-supervised learning to provide reliable predictions.
- Describe a prospective protocol for the application of machine learning in daily clinical practice.
Datasets
The project will make use of both publicly available datasets, such as LiTS (Liver Tumor Segmentation Challenge), data available through our collaborators, and a regional rectal cancer cohort from the University Hospital of North Norway (UNN). On average, 80 patients per year are operated at UNN. The regional rectal cancer cohort consists of patients’ data from 2006 to 2011, which is manually extracted from patient electronic medical journals (DIPS). It contains 376 patients who were treated surgically for rectal cancer with curative intent and monitored for liver metastases in UNN. The regional rectal cancer cohort is unique compared to other similar studies carried out by other research groups, as the inclusion criteria (e.g., patients with rectal cancer) were defined prospectively. Also, the cohort includes patients who have had rectal cancer, but no metastases in the liver, i.e., the database also includes negative cases.
Interdisciplinary Research Group
The project is interdisciplinary and performed in collaboration with UNN and UiT. The project is led by Kim Erlend Mortensen MD Ph.D., and Eirik Kjus Aahlin MD Ph.D. from the Digestive surgery research group at UiT/UNN. Prof. Robert Jenssen, Karl Øyvind Mikalsen MSc Ph.D., and Michael Kamffmeyer MSc Ph.D. are close collaborators and mentors from UiT Machine Learning Group, the group’s new Centre for Research-based Innovation (SFI) called Visual Intelligence and The Norwegian center for clinical Artificial intelligence (SPKI). Currently, Ph.D. candidate Keyur Radiya (UNN/UiT), Eirik Agnalt (UiT – machine learning group), and SPKI staff are working on the project.
