When algorithms read medical images

AI in medicine: danger or opportunity? The heated debates surrounding artificial intelligence, which can replace workers, should also shed light on its advantages, particularly in terms of healthcare and personalized medicine. Certain tasks that are too time-consuming and human-intensive to be financially viable can be delegated to algorithms. This is particularly true in the field of precision medicine, based on images acquired during clinical examinations. In fact, these images, intended for a specific purpose, are also used to train self-learning computer models, which require large quantities of data. They reveal the diversity of human organisms and the pathologies that affect patients.

CT scanners: when the technology that heals radiates into the body.

The research work of Professor Adrien Depeursinge's team has focused on images taken from CT (computed tomography) scanners. If repeated too often, this type of medical imaging poses a danger to patients, whose bodies absorb X-rays and thus a dose of radiation with carcinogenic potential. What's more, the images produced vary according to the scanner used and the type of settings requested by the radiologist. The images analyzed are therefore heterogeneous, which complicates the task of the algorithmic learning model. The "TCIA data QA4IQI" (Quality assessment for interoperable quantitative CT-Imaging) project aims to measure and isolate these variations in images, to make them homogeneous and erase differences generated by acquisition protocol, quality, radiologist preferences or software updates.

To safeguard patients' health, the team of researchers at the Institute of Informatics of the HES-SO Valais-Wallis collaborated with start-up PhantomX, which produced a 3D physical object. This realistic reproduction of a human body, or phantom, is used to calibrate the scanner and check variations produced by the machine according to settings or the X-ray dose chosen. The phantom has the advantage of remaining immobile, unaffected by radiation and able to move rapidly around different hospitals. Radiologists can thus check the variability of scanner images and harmonize them to improve readability.

SPHN funds personalized medicine research and promotes inter-institutional exchanges.

This research project, funded by the SPHN (Swiss Personalized Health Network), is being carried out in collaboration with Professor Bram Stieltjes of Basel University Hospital, Professor Ender Konukoglu of ETH Zurich, and Professor Henning Müller of the Institute of Informatics. The image acquisition and imaging physics skills of the Basel hospital and the computer vision expertise of ETH enabled the Institute of Informatics’ team to create the 3D phantom, select the relevant images, test the synthetic regions to be analyzed, check the stability of the model, and select, evaluate, and put online a dataset for the entire scientific community. Assessing the quality of quantitative, interoperable imaging is a cornerstone of personalized medicine, which is why the SPHN has supported this project. It is even more interesting that the results obtained are now being made available to the global scientific community.

Improving patient health and advancing research through open science.

The research institutes of the HES-SO Valais-Wallis are particularly keen to make scientific research accessible to all through Open Science. This way of doing research and disseminating its results enables openness, collaboration, and transparency, so that research can move forward more quickly and efficiently. Sharing data in this way required a great deal of effort on the part of Roger Schaer, scientific collaborator, and Oscar Jimenez Del Toro, former scientific collaborator at the Institute of Informatics, who spent hundreds of hours sorting, formatting, and labeling complex data; however, this is particularly important in the field of eHealth, as these technological advances directly benefit people whose health is affected. The interest shown by the scientific community when this dataset was put online, and during the summer university in which Professor Depeursinge took part, was evident.

Freeing up doctors' time to focus on patients.

The SPHN funding aims to harmonize Swiss personalized medicine practices, and the effort to standardize the CT scanner images processed by this research project will enable to verify the robustness of the artificial intelligence models already in use in hospitals. Using medical informatics developed at the Swiss Digital Center in Sierre, the possibility of reducing the radiation dose while maintaining scanner image legibility is being studied. Further research should ascertain whether algorithmic reading could make radiologists' work more robust, helping them to make the necessary medical decisions and, above all, freeing up precious time for patient relations.

Scientific publication: Jimenez-del-Toro, Oscar MD, PhD∗; Aberle, Christoph PhD†; Bach, Michael MD†; Schaer, Roger BSc∗; Obmann, Markus M. MD†; Flouris, Kyriakos PhD‡; Konukoglu, Ender PhD‡; Stieltjes, Bram MD, PhD†; Müller, Henning PhD∗,§; Depeursinge, Adrien PhD∗,∥. The Discriminative Power and Stability of Radiomics Features with Computed Tomography Variations: Task-Based Analysis in an Anthropomorphic 3D-Printed CT Phantom. Investigative Radiology 56(12):p 820-825, December 2021. | DOI: 10.1097/RLI.0000000000000795, https://journals.lww.com/investigativeradiology/toc/2021/12000

Dataset: Task-Based Anthropomorphic CT Phantom for Radiomics Stability and Discriminatory Power Analyses (CT-Phantom4Radiomics), Images & Segmentations (DICOM, 42.5 GB), https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=140312704