As every year, but in a virtual format, radiologists across the globe gathered to learn about state-of-the-art practices, to discuss case-based reviews and to present their research. This year, the context of the pandemic made sharing their experiences and discussing findings even more important. Across the talks and educational sessions, all modalities of imaging were represented, ethical issues were debated and the whole range of clinical applications - from the brain (neurology) down to the toes (traumatology) - was covered. I have chosen to highlight six exciting new studies, presented during the scientific sessions, that illustrate this diversity.
Functional connectivity of gliomas is associated with patient survival
Giulia Sprugnoli, MD, from the University Hospital of Parma (Parma, Italy) in collaboration with Dr Emiliano Santarnecchi from Beth Israel Deaconess Medical Center (Harvard Medical School, Boston, MA, USA) and Prof. Alexandra Golby from Brigham and Women's Hospital (Harvard Medical School, Boston, MA, USA) described the integration of gliomas into brain circuitry and how it can be used as an index of overall survival. It has been recently shown by several teams in human-derived xenografted mice that the integration of glioma in the brain is associated with progression of the tumour and a lower survival rate of the animals.
In the presented work, Dr Sprugnoli and her colleagues used resting-state functional magnetic resonance imaging (rs-fMRI) of patients to confirm those preliminary findings. The lesions of 54 patients with newly diagnosed (n = 18) or recurrent glioma (n = 36) were manually segmented and rs-fMRI was analysed to quantify the functional connectivity of the solid tumour with respect to the brain. A regression model between the brain integration, reflected in the functional connectivity, and overall survival with age and gender as covariates was performed. Significant functional connectivity between solid tumour and healthy brain structures were observed and resembled well-known resting state networks. Moreover, the integration profile of solid tumour predicted overall survival both in patients with newly diagnosed and recurrent high-grade glioma, outperforming classical predictors such as genetic (MGMT methylation status, IDH status), clinical (Karnofsky performance status score, tumour volume) and demographic factors.
Even though confirmation in an independent cohort of patients is warranted, these in vivo findings reinforce the hypothesis that integration of gliomas into brain networks might be related to glioma aggressiveness. It would also be interesting to correlate these findings with cognitive data of the patients or to explore the biological bases of this tumour integration. Nevertheless, this study could serve as the basis to develop a prediction model of glioma progression. It could also pave the way to new therapeutic approaches as inhibiting the communication of the tumour with the brain might improve survival. Supplementing the current regimens of drug-based and radiation treatments, Dr Sprugnoli and her colleagues envision the use of non-invasive brain stimulation to modulate the activity of the tumour and of the connected brain regions to hamper the functional connection and hopefully improve survival rate.
Eliminating gadolinium contrast agent for some breast MRI exams
Sarah Eskreis-Winkler, MD, from Memorial Sloan Kettering Cancer Center (New York, USA) evaluated the possibility to change the MRI imaging protocol used for the follow-up of probably benign breast lesions without affecting the management recommendations for the patient. Gadolinium, commonly used as a contrast agent for MRI exams, has been shown to accumulate in the brain and the long-term effects on health are still unknown. Following the fundamental principle of exposing the patient to as low a risk as reasonably achievable, Dr Eskreis-Winkler and her colleagues assessed the potential of non-contrast-enhanced breast MRI (T2-weighted images) to provide the same information as gadolinium-based contrast images.
The team retrospectively reviewed the 266 breast MRIs that had been classified as probably benign in their centre in 2017. 54 (20%) of those images presented a T2 correlate and were subsequently analysed. Three experts independently reviewed the initial exam, in which the probably benign status of the lesion was first determined, and the T2-weighted images from the 6-month follow-up. They were asked to record whether the lesion has increased and to assign the lesion to a group: probably benign (recommending another short-term imaging follow-up) or suspicious (recommending biopsy). Two weeks after this initial evaluation, the three readers were now given access to the full MRI exam from the 6-month follow-up (including the DCE-MRI sequence with contrast agent) and were asked to record the lesion change and perform the assignment again. They found an excellent agreement between the change in lesion size observed on T2-weighted and DCE-MRI images, thus confirming their initial hypothesis. The management recommendations made using the images with or without contrast agent were the same in more than 98% of the cases.
This study suggests that gadolinium-based contrast agents might be safely eliminated from these specific exams without affecting the recommendations for management. Adopting this protocol would benefit the patient as it would not only reduce deposition of the contrast agent in the brain deposition, but also reduce the use of intravenous catheters and the overall scan time. The promising results of this initial study need to be confirmed in a larger retrospective study, currently underway. In order to ultimately integrate this new method into clinical practice, prospective trials evidencing non-inferiority of non-contrast-enhanced MRI are warranted.
Surpassing traditional breast cancer risk predictor models using only images and deep-learning
Leslie Lamb, MD, from the Breast Imaging Division of the Department of Radiology of Massachusetts General Hospital (Boston, MA, USA) presented a novel deep learning model based on imaging biomarkers alone to predict future breast cancer risk. The performance of traditional commercial risk models is based on only a small fraction of patient data available from questionnaires such as age, family history of breast cancer, and hormonal and reproductive history. Risk models have only incorporated image-based features recently, when the Tyrer-Cuzick (TC) model added mammographic breast density. Yet, there is a wealth of information embedded in a mammogram beyond density that is not captured by current risk models.
Dr Lamb and her colleagues from Massachusetts General Hospital and Massachusetts Institute of Technology, proposed to leverage deep learning to extract biomarkers predicting patient's risk of cancer from mammographic images alone. The deep learning risk assessment model was developed using retrospective patient data from their centre. Nearly 250,000 mammograms in approximately 80,000 patients were used for training, testing, and validation. A traditional risk model, TC version 8 (TC8), was also applied to the same patients and the performances of the two risk prediction models were compared. The AUC for their deep learning model was 0.71 compared to an AUC of 0.61 for the TC8 model, indicating that mammograms contain highly predictive biomarkers of future cancer risk, not identified by traditional risk models.
Deep learning models are sometimes criticised for reinforcing existing biases against minorities. However, this study showed that carefully designed AI models can overcome biases existing in current commercialized traditional models. Across all sub populations, including race, the AUCs of the deep learning model were improved as compared to the TC8 model. It should, however, be noted that 81% of the enrolled patients were white and a more diverse population should be included. The deep learning model has been externally validated at centres in Sweden and Taiwan. Dr. Lamb and colleagues are actively involved in further validation studies in larger African American and minority subgroups.
A fully automated method to measure body composition from routine CT images
Kirti Magudia, MD PhD, from the Department of Radiology & Biomedical Imaging of UCSF (San Francisco, CA, USA) described a deep learning approach to evaluate body composition from routine abdominal CT images and its application to the prediction of cardiovascular events in a large cohort of patients. Two persons of similar height and weight (same BMI) but with a different ratio of subcutaneous fat to skeletal muscle can present strongly different cardiovascular risks. Indeed, subcutaneous, and visceral fat are known to increase the cardiovascular risk through insulin resistance, remodelling of the left ventricle of the heart, lowering the cardiac output and increasing systemic vascular resistance. However, measuring body composition from CT images requires manual segmentation and is too costly to be performed in routine clinical practice.
Dr Magudia and her colleagues developed a fully automated algorithm that identifies the axial series from complete abdominal CT exams found in the Picture Archiving and Communication System, and then automatically selects the slice located at the L3 vertebra level, and finally segments the image for skeletal muscle, visceral fat and subcutaneous fat with a 2.5% failure rate. A cohort of 12,128 patients, without cardiovascular disease or cancer, was followed for 5 years after an index CT exam. 1560 myocardial infarctions and 938 strokes occurred for patients in this cohort. Body composition z-scores were calculated for each patient of the cohort; the patients were divided into four quartiles for each body composition parameter; finally, a multivariate time to event analysis was performed with all body composition parameters, as well as weight, height and usual cardiovascular risk factors (smoking status, diabetes etc). Normalized visceral fat area was found to be associated with future myocardial infarction and future stroke: the patients in the 4th quartile, with the highest fat content, presented a higher risk of myocardial infarction and the patients in the 1st quartile, with the lowest fat content, were protected from stroke. Therefore, this study confirms the potential of this automated method and paves the way for additional large-scale studies about the role of fully automated CT-based body composition analysis in the management of cardiovascular and oncologic diseases.
Epicardial fat and coronary atherosclerosis in the HIV population
Manel Sadouni, MD, from the Centre de Recherche du CHUM (Montreal, Canada) discussed the results of a cross-sectional study about lipodystrophy in patients with HIV and its association with coronary atherosclerosis. HIV-infected patients develop changes in fat distribution that are associated with an increased risk of cardiovascular diseases and as a result, coronary atherosclerosis is a major cause of death in those patients. Dr Sadouni and her colleagues previously evidenced that the volume of epicardial fat was increased in patients with HIV compared to controls. Antiretroviral therapy was also found to play a role in the fat increase and the volume of epicardial fat was associated with an increase in non-calcified plaque volume.
In the present study, Dr Sadouni investigated the attenuation of epicardial fat on CT images, i.e. the focus of the study was not on the quantity of fat but on its quality as CT attenuation is a potential surrogate of the inflammatory activity of the fat. This sub-study is nested in the large Canadian HIV and Aging Cohort study involving 10 centers in Canada and following more than 1000 HIV+ and HIV- patients. 263 patients met the inclusion criteria for this study (179 HIV-infected patients and 84 healthy controls) and were invited to undergo cardiac CT. Assessment of volume and CT attenuation of epicardial fat, as well as volume of total atherosclerotic plaque were performed by two observers blinded to HIV status and clinical history. The two groups of patients presented the same cardiovascular risk, measured by the 10-year Framingham score. Epicardial fat volume and epicardial fat attenuation index (EFAi) were observed to be higher in HIV+ patients compared to HIV- participants. Furthermore, using multivariate analysis, EFAi was associated to HDL-cholesterol, BMI and to duration of antiretroviral therapy. After adjustment for cardiovascular risk factors, EFAi was significantly associated with total coronary plaque volume.
The cross-sectional nature of the study does not allow to infer causality but the observed association of epicardial fat CT attenuation with antiretroviral therapy duration and subclinical coronary artery plaque suggest a potential mechanism that could explain the increased risk for coronary artery disease in the HIV population.
The PATCH trial: treatment of cerebrospinal fluid leaks in spontaneous intracranial hypotension
Timothy Amrhein, MD, from the Department of Radiology of Duke University (Durham, NC, USA), reported the results of an initial sham clinical trial, called PATCH, evaluating the feasibility of randomization and blinding of the first line treatment for spontaneous intracranial hypotension (SIH), epidural blood patching. SIH is a debilitating condition caused by spontaneous spinal cerebrospinal fluid leaks and that presents with orthostatic headache and a myriad of other cranial nerve-related symptoms (dizziness, nausea, cognitive dysfunction, hearing loss etc.). Patients can be treated conservatively (with bed rest, hydration etc.), with blood patching or with surgery. Imaging-guided epidural blood and fibrin glue patching is accepted as the optimal therapy but the efficacy of this procedure has not been demonstrated in a randomized controlled trial.
Dr Amrhein and his colleagues designed a single-centre, parallel, blinded, controlled trial where participants were randomly assigned to CT fluoroscopy-guided targeted blood and fibrin glue patching, or to a simulated procedure with saline injection instead of patching. The primary outcome measure was the Headache impact test-6 (HIT-6) at 1 month. The changes in SIH signs from baseline to two months after the procedure were assessed with the Dobrocky score from brain MRI. Among the 83 patients who were eligible for the procedure, 68 declined to participate. 15 patients were thus enrolled. Blinding was successful as 47% participants did not know their intervention arm and 75% of the participants who thought they knew were actually wrong. A reduction was observed in the HIT-6 and Dobrocky scores indicating a trend towards improvement of headaches, even though this trend was not statistically significant, presumably because of the small sample size. The successful randomization and blinding observed in this trial confirms the feasibility of conducting a definitive multi-centre randomized controlled trial to evaluate the efficacy of this therapy in the SIH population.
Conclusion
Overall, this meeting made clear that radiology is leading the way with artificial intelligence in medicine. Important technical advances were presented and debates about ethical or practical considerations for translation of those techniques to the clinic were organized. Even though we could not discuss all the related presentations for sake of concision, we are looking forward helping to disseminate research on those topics in the coming months.