ESR Research Seed Grants 2022

Artificial Intelligence

• The Brain Age Biomaker Initiative in Stroke 

Summary: We built a dedicated artificial intelligence framework to derive a personalized brain health MRI biomarker from a sizeable institutional cohort of stroke patients treated by mechanical thrombectomy for an acute ischemic stroke due to a large intracranial arterial occlusion. Leveraging MRI from 371 thrombectomy patients, we predicted Relative Brain Age from admission T2-FLAIR MRI. This biomarker expresses quantitatively if a patient has an older- or a younger-appearing brain on MRI. We then showed that patients with a history of hypertension or diabetes mellitus had an older-appearing brain. Finally, we showed that patients with older-looking brains had poorer outcomes after stroke, regardless of stroke severity, chronological age, or treatment modalities. We believe our biomarker provides a preliminary platform for the conceptualization of personalized stroke care as it can identify patients with brain frailty upon admission with an easily understandable biomarker.

Martin Bretzner Department of Neuroradiology, Lille University Hospital, Lille/FR

• PI-QUAL and artificial intelligence: a pilot study to test a novel tool to assess the quality of multiparametric MRI of the prostate

Summary: Image quality plays a fundamental role in multiparametric magnetic resonance imaging (MRI) of the prostate (that includes three different sequences, T2-WI, DWI and DCE) to safely rule in and rule out clinically significant prostate cancer. The Prostate Imaging Quality (PI-QUAL) score is the first standardised scoring system that outlines specific
quality criteria for the acquisition of prostate MRI and we previously built a semi-automated tool to assess image quality using this score. With this pilot study we wanted to explore if our tool has the potential to become fully automated and also
to shed some light on the future steps that need to be taken with the aid of artificial intelligence. We tested our tool both in a training and in a validation set (for a total of 151 MR scans from different centres) and found promising results that suggest that it is possible to evaluate the diagnostic quality of each single MR sequence in prostate MRI automatically. However, at present we still need the human (i.e., the Radiologist’s) subjective assessment to give an overall image quality score using PI-QUAL and we plan to conduct future research with the aid of artificial intelligence to fine tune our tool.. 

Francesco Giganti Radiology/UCL Division of Surgery and Interventional Science, University College London Hospital, London/UK

• Diffusion Probabilistic Models to Reduce the Use of Contrast Agent in Breast MRI

Summary: Current EUSOBI guidelines recommend MRI-based breast cancer screening for women with dense breasts.
Breast MRI requires a contrast-enhanced dynamic sequence for the diagnosis of lesions, and thus the administration of a gadolinium-based contrast agent (GBCA). In the screening context, a reduction of GBCA dose is favorable. However, this results in subtraction images with reduced contrast-to-noise ratio (CNR). The goal of this project was to employ denoising diffusion probabilistic models (DDPM) for generating high-dose from low-dose images, i.e., increasing CNR. DDPM are particularly suitable, because they learn to denoise images by adding noise to training images in progressive amounts and reverting this process step-by-step. After setting up the DDPM, noisy input images at 25%, 10% and 5% of the standard dose were calculated from original subtraction images. These were fed into the DDPM, and, for comparison purposes, also into generative adversarial models (GAN). Two breast radiologists rated the generated high-dose images. SSIM as numerical image metric was calculated between original and generated images. Both radiologists preferred the DDPM-generated images at 25% input dose, but preferred the GAN-generated ones at 5% (all P<0.05). Lesion conspicuity was rated equivalent between GAN and DDPM (all P>0.05), except for 5%, for which the GAN yielded significantly better lesion conspicuity (P<0.001 and P<0.05 for reader 1 and 2, respectively). SSIM did not differ between models (P>0.99 for all doses). In conclusion, DDPMs are highly promising for the task of denoising low-dose subtraction images as they can cope with different input noise levels. 

Teresa Nolte Diagnostic and Interventional Radiology, University Hospital RWTH Aachen, Aachen/DE

• Fetal MRI radiomics for machine learning-assisted lung maturation assessment and outcome prediction

Summary: This project investigated whether radiomics, i.e. a large number of automatically extracted image features with a focus on shape and texture, may be useful to gain information from fetal MRI data. For this purpose, radiomics features extracted from the fetal lung in repeated fetal MRI acquisitions were compared. It was shown that a majority of features describing the shape and texture of the developing lung are robust with regard to repeated acquisitions. Thereby, this project highlights that through the use of radiomics, fetal MRI is able to deliver reliable image features that, while not visually accessible, may be helpful in characterizing fetal lung development. In a limited number of cases this project demonstrated possible future applications of radiomics in fetal MRI: It was shown that radiomics features could be used to estimate fetal gestational age based on fetal lung images. Further, radiomics could differentiate between a healthy control group and age-matched fetuses with pathologic lung development, specifically congenital diaphragmatic hernia. In the future, radiomics may be used to complement human reader assessment of fetal MRI data and allow reliable characterization of normal and pathologic fetal development and beyond.

Publications:
Prayer, F. et al. Fetal MRI radiomics: non-invasive and reproducible quantification of human lung maturity. Eur Radiol 33, 4205–4213 (2023). doi: 10.1007/s00330-022-09367-1
Watzenboeck, M.L. et al. Reproducibility of 2D versus 3D radiomics for quantitative assessment of fetal lung development: a retrospective fetal MRI study. Insights Imaging 14, 31 (2023). doi: 10.1186/s13244-023-01376-y

Florian Prayer Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna/AT

• Artificial Intelligence for Fracture Detection in Paediatric Osteogenesis Imperfecta

Summary: Osteogenesis imperfecta (OI) is a genetic condition characterized by skeletal fragility and increased risk of fractures. Most patients suffer from long term disability and pain from their fractures, which are frequently missed due to the unusual appearances of their bones and lack of experience from healthcare providers of this relatively rare disease (1 in 10,000 affected). This study evaluated whether the use of a commercial AI tool for fracture detection (not trained on children or those with OI) could still help radiologists in detecting fractures in children with OI, or whether the use of such a tool may place this population at risk of harm if AI were more widely implemented for routine usage. We evaluated the difference in diagnostic accuracy across 7 radiologists (all with a paediatric radiology interest) and found that even in this ‘expert’ group, AI assistance could demonstrate a small increase in accuracy. It is unclear whether this improvement would be clinically significant in the ‘real world’, however it provides some reassurance that the AI tool did lead to any reduced reader accuracy and, when used together with a radiologist, it may not necessarily cause patient harm even though it was not specifically trained to review fractures in this niche population.

Susan Shelmerdine Clinical Radiology, Great Ormond Street Hospital for Children NHS Foundation Trust, London/UK

• A Convolutional Neural Network for Automated Segmentation of Solid Renal Tumors on CT Images

Summary: This project aims to develop an AI-algorithm for automated segmentation of solid renal tumors on CTimages with comprehensive visualization. Therefore, we implemented a deep-learning approach (DeepLab v3+ segmentation network with ResNet 50 backbone using pytorch). Renal tumor segmentation was independently performed by an experienced GUradiologists and two research assistants. Predictions of separate reader-based AI-algorithms were combined using so-called ensemble methods. AI-algorithms were trained and validated (80% / 20% split) on a database containing n=639 renal tumor patients (median age 66 years; 64% male; median tumor diameter 47 mm) imaged in corticomedullary and nephrogenic contrast media phase (n=1077 separate CT scans). The CNN ensemble method achieved a median DICE-Score=0.82/0.84 in the internal validation data for corticomedullary / nephrogenic CT scans, respectively. All renal tumors were as least partially identified (100%). Segmentation predictions were visualized using contour- and color-coded maps with overlay on
clinical CT scans. Segmentation thresholds scan be individually chosen by radiologists to optimize result during clinical application. Testing was performed on an independent external dataset (2019 TCIA KITS data) containing n=210 patients with renal tumors imaged with corticomedullary phase CT. Here, the AI-algorithm identified 99% of renal tumors at least partially and achieved a median DICE-Score=0.80. Thus, the here presented AI algorithm provides a robust automatic renal tumor segmentation that is generalizable to external CT data. Segmentation predictions can be easily visualized in a clinical imaging environment using contour- and color-coded overlay maps.

Johannes Uhlig Department of Diagnostic and Interventional Radiology, University Medical Center Goettingen, Goettingen/DE