IHI project IMAGIO developed a tool that evaluates whether a treatment that uses heat to destroy cancer cells has successfully eliminated the tumour.
For many years, people diagnosed with a type of liver cancer called hepatocellular carcinoma had to undergo surgery to cut the tumour out. Nowadays, one well-established treatment option for (very) early-stage tumours is a minimally-invasive technique called thermal ablation. Here, a specialist called an interventional radiologist uses imaging techniques to guide a needle-like probe directly into the tumour.
Once in place, the probe is heated and the heat destroys the cancer cells. Thermal ablation is an effective treatment and has a number of benefits over surgery, including quicker recovery times, fewer complications, and a shorter stay in hospital.
A key question: has the tumour been fully destroyed?
The challenge for doctors is to check if the tumour has been fully destroyed, or if some cancerous tissue has escaped the treatment. This task entails evaluating the ‘ablative margin’ – the small rim of healthy tissue around the tumour. The check is important because it will determine if the patient needs additional treatment or not.
One option is to do this manually; here doctors assess the scans by eye and mark them manually, something that requires significant amounts of both time and expertise. Automated tools have been developed to carry out the task, but are not always accurate, particularly when image quality is poor.
Now, a team of researchers from the IHI project IMAGIO has developed a new artificial intelligence (AI) tool called AblationNet. The tool is described in a paper published in the journal Biomedical Signal Processing and Control.
A deep-learning model trained on a multicentre dataset
The researchers’ first step was to study the biggest challenges involved in identifying ablation zones on computed tomography (CT) scans, such as handling wide variations in image quality.
Based on these challenges, we designed a deep-learning model that can learn both the overall shape and detailed boundaries of the ablation zone at different spatial scales. The model was trained and tested using expert annotations from a multicentre dataset. This multicentre approach was intended to improve the robustness and generalisability of the model to the variability encountered in clinical CT images.
Faeze Gholamiankhah, the first author of the paper, a PhD student at the Department of Radiology at Leiden University Medical Centre (LUMC) in the Netherlands.
The public-private nature of the project allowed the team to easily gather input from both academic and industrial partners, identify knowledge gaps, and gain insights into similar products under development or on the market.
A strong performance, even when image quality is low
Tests show that AblationNet consistently performs better than existing tools, even in cases where the images are of low quality. Furthermore, the researchers point out that AblationNet performs the task in one single step, eliminating the need for user input or manual correction. The team now plans to test the models further.
‘The models developed within the IMAGIO project will be implemented in software developed in-house at LUMC, which is currently used to analyse CT images acquired during thermal ablation procedures,’ said Ms Gholamiankhah. ‘This will enable us to further evaluate the models within our existing analysis workflow and facilitate their potential translation into clinical practice.’
The work on AblationNet will also feed into IMAGIO’s efforts to develop other AI-based models, such as one to predict the risk of tumour recurrence following liver thermal ablation.
Looking to the future, the IMAGIO team hopes that their tool will ultimately support doctors during their evaluations of thermal ablation procedures, helping them to make better-informed decisions about the next steps for their patients.
Meanwhile patients who undergo thermal ablation will benefit from improved care and greater confidence in the outcomes of the procedure.
