Medicine and health

Medicine and health

We develop more efficient deep learning methods for diagnosis support and decision support for diseases such as cardiovascular diseases and cancer.

Motivation

In recent years, research on deep learning has contributed to significant progress for aiding diagnostic and decision support, treatment planning and healthcare resource allocation. This is primarily done by extracting relevant information from medical data, as well as clinical datasources such as time series measurements or radiology reports.

However, analyzing medical data is achallenging and time-consuming task. Our innovations in medicine and health domain aim to create more efficient diagnostic and decision support tools for diseases like cardiovascular diseases and cancer by using deep learning.

By developing well-performing deep learning models in the medical domain, we assist healthcare professionals by increasing certainty and streamlining analyses of medical images.

Our innovations

Our deep learning research in the medical domain has contributed to several innovations which aim to assist healthcare professionals in the clinical workflow, including methods for:

• automatically measuring the heart's left ventricle in 2D echocardiography, in collaboration with GE Vingmed Ultrasound.

By making the system more aware of the heart's structure, it can measure the left ventricle more accurately and consistently, providing quicker results for patients and less workload for doctors. Read more.

• predicting future breast cancer risk based by analyzing mammography images, together with the Cancer Registry of Norway.

This method demonstrates consistent improvements in predicting breast cancer risk up to five years in advance, paving the way for more personalized screening, earlier identification of high risk patients, and more efficient use of screening resources.

• estimating tumor-infiltrating lymphocytes (TIL) in non-small cell lung cancer, in collaboration with the University Hospital of Northern Norway (UNN).

The model represents a resource-efficient deep learning pipeline for estimating TIL density in this lung cancer type, which could lead to faster treatment of lung cancer, less costly examinations, and more personalized cancer treatment. Read more.

• estimating the arterial input function in dynamic PET scans, in collaboration with the University Hospital of Northern Norway (UNN).

This model estimates the AIF directly from PET data, providing a non-invasive and cost-effective alternative to arterial blood sampling. Read more.

• better augmenting CT images by clipping intensity values tailored to characteristics of organs, such as the liver, in collaboration with UNN.

The model proposes a new approach to augment CT images with the purpose of detecting lesions in the liver, and is currently being incorporated for prospective studies.

• content-based image retrieval of CT liver images using self-supervised learning, together with UNN.

This system automatically retrieves CT images of the liver from a large database, without the need for manual and time-consuming searches. Read more.

Addressing research challenges

Developing deep learning methods for medical purposes involve several research challenges, such as the availability of training data, estimating the models’ confidence and uncertainty, as well as their lack of explainability and reliability. The aforementioned innovations address these challenges in multiple ways.

For instance, the methods’ ability to learn from limited data is at the core of the data augmentation technique for CT images. The method also leverages context and dependicies by exploiting knowledge about the signal-generating process.

Explainability and reliability constitute a significant part of our method for detecting cancer in mammography images, as well as our content-based CT image retrieval method.

Synergies within the innovation area and across other areas

When developing deep learning solutions for medical and health-related challenges that our user partners face, it is important to transfer knowledge and methodologies both within and across innovation areas. Our proposed medical methods synergize well with our other three innovation areas.

Our semi-automatic landmark prediction model depends on context provided by echocardiography scan lines. This is inspired by centre-developed methods which leverage anatomical knowledge, such as for detecting cancer in mammography images.

Self-supervised deep learning, which is a staple of our medical innovations, has also proven useful for addressing research challenges in marine, energy, and earth observation domains. For example, our CT image retrieval framework shares similarities with a content-based image retrieval system for seismic data.