AI Risk Score Predicts Breast Cancer Risk Better Than Mammographic Density

January 2, 2020

The latest roundup of radiology news and studies.

AI Risk Score Outperforms Mammographic Density for Breast Cancer Prediction

Today, radiologists rely on mammographic density assessments, in conjunction with questionnaires and other clinical information, to create risk models for breast cancer. Researchers from the Karolinska Institute have now published a study in Radiology that suggests a deep learning artificial intelligence (AI) model, trained on mammographic images, can provide a more accurate way to predict who will go on to develop the disease.

To compare traditional methods of assessing risk with a deep learning network, the study authors evaluated 2,283 women, 278 of which had been diagnosed with breast cancer, in a retrospective study. The network was developed using images from cases diagnosed over a four-year period, using the earliest available mammography data for each woman. The model produced a deep learning risk score, as well as dense area metrics and percentage density. When the model was, then, tested using mammographic images from both women who had breast cancer and healthy controls, they discovered:

• Age at mammography, dense area, and percentage density were higher in women who were diagnosed with breast cancer;
• The odds ratios and areas under the receiver operating characteristic curve were higher for age-adjusted DL-risk scores;
• and the use of the model resulted in few false-negative results, particularly in more aggressive forms of the disease.

Taken together, the authors concluded that, compared with density-based models of risk assessment, the use of AI algorithms, deep learning networks in particular, may help clinicians better predict which women may be at highest risk for developing breast cancer in the future.

CT Radiomic Features of Pancreas Can Aid in Diagnosing Abdominal Pain

A new study, published in the European Journal of Radiology, demonstrates that quantitative radiomic features of the pancreas in CT scans can help radiologists differentiate which patients have functional abdominal pain, recurrent acute pancreatitis, or chronic pancreatitis.

When a patient presents in the emergency room with strong abdominal pain, it can be difficult to assess what the clinical issue may be. Traditionally, when the patient says the pain is recurrent, and pancreatic enzymes are at elevated levels, doctors may diagnose those patients with recurrent acute pancreatitis, regardless of any imaging studies being done.

Researchers from the Johns Hopkins University, however, wondered if radiologic features, present on contrast-enhanced CT scans, could help physicians provide more accurate diagnoses. To test the idea, the study authors did a retrospective analysis of 56 patients who had been seen in the pancreatitis clinic with recurrent acute pancreatitis, functional abdominal pain, or chronic pancreatitis. After outlining the pancreas on images, the authors were able to extract 54 distinct radiomic features to compare between the patient groups. They also relied on a one vs. one Isomap and Support Vecture Machine (IsoSVM) classifier to help sort the groups based on radiomic features.

Careful analysis revealed that 11 radiomic features were significantly different between the three patient groups – and using those features, the ISOSVM classifier was able to predict the appropriate diagnosis with an 82.1 percent accuracy rate. The results led the authors to conclude that such radiomic features can help radiologists provide more accurate diagnoses of the cause of abdominal pain with CT scans in the future.

Combined Intelligence: Radiologists and AI Improve Pneumonia Diagnosis

A new study published in npj Digital Medicine suggests the combination of a radiologist and an artificial intelligence (AI) platform can better diagnose pneumonia from chest radiographs.

Too often, the rise of AI is framed as a doctor vs. machine type of debate. Certainly, most studies published, to date, have directly compared the performance of seasoned radiologists with these deep learning algorithms. Yet, new studies are evaluating ways to measure how AI systems may augment radiologists’ abilities – and prove to be more accurate and efficient than either doctor or algorithm alone.

Researchers from Stanford University and Duke University developed a “human in the loop” method of AI design, which uses a combination of automated detection and human checks to identify pneumonia on standard chest X-rays. When they placed this AI system on Swarm, a platform that allows multiple radiologists to work together, the accuracy of their diagnoses significantly increased.

The researchers recruited 13 experienced radiologists, split them into two groups, and had them estimate the likelihood of pneumonia on a data set of 50 chest radiographs. They did so alone and, then, again using the Swarm platform with embedded AI. In addition, the researchers utilized two deep-learning models, CheXNet and CheXMax, to review the images. When the researchers compared the radiologists alone, the AI alone, and then a combination of real-time Swarm platform, the researchers discovered that the combination of AI and human experts provided a “symbiotic” effect, outperforming any one method used in isolation. The study authors concluded that finding ways to use AI to augment, rather than replace, the abilities of radiologic experts will help improve diagnostics in the future.

An Assessment of Ethics in Radiology

The practice of medicine is fraught with ethical dilemma, including end-of-life decisions and appropriate doctor-patient interactions. While radiologists, historically, may not be involved in these kinds of situations, there are still many ethical issues to consider in this specialty. Yet, past studies have found that many radiology residency programs do not offer ethical education or seminars for trainees. Now, researchers from Penn State Health Milton S. Hershey Medical Center, Johns Hopkins Medical Institutions, University of Pennsylvania, and University of Washington recently published the results of a questionnaire regarding common perceptions and experiences regarding ethical concerns in the American Journal of Roentgenology.

The study authors recruited 424 radiologists and trainees through email invitations, social media postings, and requests for participation on radiology-related websites and web blogs. Fifty-six percent of those respondents stated they had witnessed an ethical dilemma, generally regarding how errors were handled. Trainees were more likely to report any mistakes, either on their part or on the part of colleagues, to the patient than more established radiologists. When it came to negligence, however, while 81 percent stated they would report a negligent act to a supervisor, practicing radiologists were more likely to say they would do so than those still in training.

The authors concluded that there are many common ethical dilemmas that radiologists face and that these dilemmas should be addressed in some manner during residency and fellowship training to provide a more “uniform” approach to managing them when encountered.

Moving Closer to Personalized Medicine in Radiology

Many cancer patients are prescribed immunotherapy to help them combat the disease. Yet, it can be difficult for radiologists to select the appropriate therapy, and then, later, to determine whether that therapy is working without invasive tissue sampling or repeatedly taking blood samples. Now, a new technique called 89Zr-immuno-PET, a whole-body molecular imaging technique, allows clinicians a way to quantify the uptake of antibodies in patients’ tissue so they can better predict the effectiveness of a particular therapy. The results were published in the Journal of Nuclear Medicine.

Researchers from the Cancer Center in the Netherlands used the technique on 36 cancer patients who were injected with 89Zr-labeled antibodies. The technique allowed them to not only visualize the tumors throughout the body but also to measure non-specific uptake of those antibodies in both normal patient tissue and tumors, including in the kidney, liver, lung, and spleen. This, the researchers argued, can help clinicians better understand what therapies will be more effective in treating cancers, as well as understand which drugs may lead to toxic effects on the body. The study authors plan to look at uptake as a function of time in a future study.