A team of researchers from Massachusetts Institute of Technology (MIT), Harvard Medical School and Chang Gung University in Taiwan has trained their deep learning model “Sybil” to use a single CT scan to assess lung cancer risk. Further clinical data were not necessary for this. They were able to make predictions about the probability of lung cancer occurring between one and six years after screening. For the future it is conceivable to use Sybil for a second opinion when evaluating lung images, as shown in the study published in the Journal of Clinical Oncology.
Created in Python and AI model released on Github under MIT license was trained retrospectively using data sets from clinical studies with smokers from the USA and data from the Linkou Chang Gung Memorial Hospital – the latter also containing data from non-smokers. Since lung cancer is also widespread among non-smokers in East Asia, LDCTs are also performed on non-smokers in Taiwan, for example. In the United States, low-dose X-ray CT (LDCT) scanning is recommended for early cancer detection in long-term smokers who have smoked for 30 years or more.
Lung cancer prediction with Sybil by years. The reported concordance index (C-index) ranges from 0 to 1. It reflects the agreement between the prediction and the observed results.
The predictions achieved a high level of agreement with the actual patient history. As expected, the accuracy of the prediction decreased the further Sybil looked into the future.
AIs have been supporting radiologists for a long time
The idea of detecting cancer from radiological data using a deep learning model is not new and is being used more and more in practice. It was only at the end of 2022 that Google announced a cooperation with the American medical technology company iCAD to support radiologists in clinical practices in evaluating 2D and 3D images. However, the AI methods are not intended to replace the doctor, since radiologists – at least in the EU – also have to look at images that the AI has marked as inconspicuous.
(mack)
