Do Physicians Want Fewer Radiologist Readings?

Survey finds physicians prefer single physician reads for multi-part CT scans

It’s often said that two eyes are better than one-but a new survey published in the Journal of the American College of Radiology suggests, when it comes to multi-part CT scans, physicians would prefer a single radiologist’s take in order to reduce potential communication failures that may lead to medical errors.

Complex imaging orders, particularly multi-part CT scans of the chest, abdomen, and pelvis, are increasing. And with so many subspecialists in the radiology field, it’s not uncommon for several radiologists to offer their take on a multi-part imaging study. Researchers from the Pritzker School of Medicine at the University of Chicago surveyed more than 200 physicians from emergency medicine, outpatient internal medicine, hospitalist medicine, general surgery, and radiology from two different institutions in order to better understand referring physician preferences as well as best practices for the communication of radiologic findings.

Related article: Clinical Roundup: The April Radiology News You Shouldn’t Miss

The survey discovered that both radiologists and referring physicians prefer to have a single radiologist read an image and communicate their findings in a single report for a multi-part CT scan to avoid “ambiguity.” This result was seen in both responses to Likert-like questions and open response fields, and the physicians showed mildly greater confidence and the possibility of “more rapid patient care decisions” when a single radiologist handled all scans.

The authors argue that it’s possible that multiple radiologist reports may lack clear interpretations or recommendations and, as such, overall, the survey results demonstrate the importance of a clear and cohesive results interpretation. Despite the growing number of radiologic subspecialties, the researchers suggest single read and report in these multi-part studies may help reduce the potential for medical errors and misdiagnosis.

Google AI outperforms radiologists on lung cancer detection

In a new study published in Nature Medicine, researchers from Google AI, Stanford Health Care, Northwestern Medicine, and New York University-Langone Medical Center demonstrated that a new deep learning algorithm, or artificial intelligence (AI) program, can better diagnose small lung cancers from low-dose CT images than trained radiologists.

Numerous studies have shown that screening can reduce the risk of lung cancer deaths-particularly in patients with a history of smoking-by identifying cancers as well as spots that may later develop into malignancies. But false positive and false negative reads do occur-which can result in patients not receiving the life-saving care they need or, alternatively, being subject to biopsies and other unnecessary procedures for a disease that doesn’t exist.

Given previous work showing that AI models can better predict certain forms of cancer from imaging data, the researchers created a layered deep learning network and trained it on thousands of CT scans. These scans included patients diagnosed with lung cancer, patients who would go on to later develop lung cancer, and patients without lung cancer. They then tested the model against 6,716 National Lung Cancer Screening trial cases whose diagnoses were already known, achieving a 94.4% accuracy rate. The algorithm performed at a similar level of accuracy on an independent data set of an additional 1,139 CT scans.

When the group compared the model to reads by six seasoned radiologists on a single CT scan without prior CT images, it outperformed the physicians, with an 11% reduction in false positive results and a 5% reduction in false negative results. When both prior and current images were available, the AI and the radiologists performed at a similar level. While this model is preliminary, the study authors argue that the development of AI programs in radiology can augment what radiologists do today, increasing the accuracy of lung cancer screening across the globe.

Texture analysis may offer new way to identify osteoporosis

Traditionally, radiologists rely on dual-energy X-ray absorptiometry (DEXA) to diagnose osteoporosis. Now, a new study published in the European Journal of Radiology suggests that texture analysis of a non-contrast CT head scan can also help radiologists discriminate between normal bone mass density (BMD) and osteoporosis.

Researchers from the Boston University School of Medicine investigated the use of texture analysis, or a form of radiomic data analysis that extracts, analyzes, and interprets quantitative imaging features, to identify which craniofacial bones might be best to sample to help identify issues with bone density. The group did a retrospective study, using non-contrast CT to image the clivus, bilateral sphenoid triangles, and mandibular condyles, extracting 41 text features from 29 patients with normal bone density and 29 patients who had been diagnosed with osteoporosis.

Of those 41 text features, the researchers identified 16 in the clivus that showed significant differences between patients with normal bone density and those with osteoporosis. Those features included four histogram features, two gray-level co-occurrence matrix features, eight gray-level run-length features, and two Law’s features. They identified 19 such significant features in the sphenoid triangles, as well as 24 features in the left sphenoid, 31 in the left condyle, and 22 texture features in both sides of the mandibular condyles.

The study authors concluded that these specific texture analysis features, particularly those found in the clivus, can help radiologists diagnosis osteoporosis in patients who have not undergone DEXA screening.

Echo-planar fluid-attenuated inversion recovery MRI offers clearer view of fetal brain

Fetal MRI is warranted when a brain abnormality, such as ventriculomegaly, is detected during a prenatal sonogram. New research suggests echo-planar fluid-attenuated inversion recovery (EPI-FLAIR) MRI can offer a better visualization of the fetal brain than traditional T2 methods, offering clearer insights into fetal brain development. The results are published in the journal Radiology.

Currently, T2-weighted single-shot fast spin-echo MRI is used to visualize the cortical plate in utero. But the method’s poor contrast between the subplate and the underlying immediate zone does not allow radiologists a clear look of other brain layers. Researchers from the Medical University of Vienna hoped that EPI-FLAIR MRI might offer a more detailed look at brain layering.

The researchers compared the delineation of fetal brain lamination using both methods and had two neuroradiologists grade the visibility of brain layers. In addition, one of the neuroradiologists performed a region-of-interest analysis in the cortical plate, subplate, and underlying intermediate zone.

A total of 259 MRI studies in fetuses with a mean gestational age of 26.9 weeks were included in the qualitative analysis while 72 examinations with a mean gestational age of 27.4 weeks were included in the quantitative analysis. The researchers demonstrated that:

• Subplate identification was superior using the EPI-FLAIR method with a higher agreement between the two neuroradiologist raters;

• Signal intensity ratios between the underlying intermediate zone and subplate were significantly higher using EPI-FLAIR;

• And subplate to cortical plate ratios were not significantly different between the two methods.

Given those results, the study authors suggest that the EPI-FLAIR method may offer neuroradiologists an improved visualization of fetal brain lamination in comparison to traditional methods, allowing for more accurate diagnoses of fetal brain abnormalities and potentially greater insights into normal fetal brain development.

Updating ICD-10 codes for head CT

According to a new study published in Current Problems in Diagnostic Radiology, the codes related to head CT scans currently listed in the World Health Organization’s 10th revision of the International Statistical Classification of Diseases and Related Health Problems are not sufficient to meet the complexity seen in patients.

Researchers from the MD Anderson Cancer Center, Massachusetts General Hospital, Thomas Jefferson University, and the Georgia Institute of Technology, among others, argued that, while other physicians can bill for different levels of evaluation and management (E&M) services based on how complex a patient case may be, radiologists are stuck using a single code no matter what the underlying medical issue may be.

Policy makers have suggested that the current head CT code may be overvalued. Since a procedure’s RVU is strongly associated with the complexity of patients undergoing the service, the study authors wanted to gauge the amount of complexity seen when a head CT is ordered.

The researchers used the 2017 Medicare PSPS Master file to identify patients who had undergone head CT examines during a visit to the emergency department. They then assessed patient complexity using the E&M level noted during that visit. They discovered:

• 56.1% of head CT exams were performed in the emergency department;

• 70% of non-contrast exams performed were ordered for patients with a level 5 E&M designation;

• The most common ICD-10 code for head CT billed for a highly complex patient encounter was “dizziness and giddiness,” even when the patient was admitted for a traumatic car accident;

• And the most common ICD-10 code for head CD without contrast was “headache.”

The authors concluded that such codes did not adequately reflect the complexity of patient cases seen in the emergency department, failing to line up with what is required by clinicians as they address level 5 patient care.

This not only potentially affects billing and reimbursement policies for radiologists but also how accurately the WHO, and other interested organizations, can track different ICD codes and related issues for research and policy direction. They recommend that further studies be done, outside of Medicare data, to better understand what codes are used and develop new codes that may better represent true patient complexity.