Journal highlights

The following are highlights from the current issues of RSNA’s peer-reviewed journals.

Radiology Logo

Generalizability and Bias in a Deep Learning Pediatric Bone Age Prediction Model Using Hand Radiographs

Although radiologists are enthusiastic about AI, deep learning (DL) algorithms have proven susceptible to pitfalls that may limit their safety and clinical readiness. Issues, including a lack of generalizability and bias, threaten the safe and equitable use of AI.

In an article published in Radiology, Elham Beheshtian, MD, University of Maryland School of Medicine, Baltimore, and colleagues examined the generalizability and bias of the winning DL model from the 2017 RSNA Pediatric Bone Age Challenge.

The model was tested from September 2021 to December 2021 on an internal validation set and an external test set of pediatric hand radiographs with diverse demographic representation. The internal validation set had images from 1,425 individuals and the external test set had images from 1,202 individuals.

According to the authors, the DL model showed similar performance for the internal validation set compared with the external test set and resulted in bone age errors in 16% of the external test set that would have led to clinically significant errors.

“A DL model trained on hand radiographs generalized well to a diverse external test set but also demonstrated clinically significant sex-, age- and sexual maturity–based biases. We recommend caution when using DL models clinically, especially if bias has not been evaluated,” the authors conclude.

Read the full article at Follow the Radiology editor on Twitter @RadiologyEditor.

Fig 3 Beheshtian

Random examples of frontal hand radiographs in healthy children without a notable clinical history from the Digital Hand Atlas data set for which there were clinically significant errors when evaluated by the 16Bit model, including (A) 33-month-old Black boy with deep learning (DL) model mean absolute difference (MAD) of 15 months (compared with ground truth), resulting in clinical diagnosis of advanced skeletal maturity (ground truth, normal); (B) 201-month-old Asian girl with DL model MAD of 24 months, resulting in clinical diagnosis of delayed skeletal maturity (ground truth, normal); (C) 189-month-old White boy with DL model MAD of 60 months, resulting in clinical diagnosis of normal skeletal maturity (ground truth, advanced); and (D) 75-month-old Hispanic girl with DL model MAD of 15 months, resulting in clinical diagnosis of normal skeletal maturity (ground truth, delayed). ©RSNA 2022


How We Got Here: The Legacy of Anti-Black Discrimination in Radiology

Current disparities in access to diagnostic imaging for Black patients and the underrepresentation of Black physicians among radiology trainees and practicing radiologists reflect contemporary consequences of a history of anti-Black discrimination in the U.S.

Although diagnostic radiology is often not a patient-facing field, it has become a key structural component of a patient’s diagnostic and therapeutic health care journey. Therefore, identifying and addressing race-based health care disparities in radiology has impacts far beyond imaging and ultimately on patient health outcomes.

In an article published in RadioGraphics, Julia E. Goldberg, MD, MBA, NYU Langone Health, New York, and colleagues researched disparities in access to imaging and health care among Black patients to illustrate how understanding radiology’s history and resultant structural racism is crucial to strive toward health equity for both patients and radiologists.

The authors noted that health care disparities negatively affect cancer-related imaging. For breast, lung and colorectal cancer, Black patients have up to a 42% higher mortality rate compared with that of white patients. In addition, structural barriers within radiology affect workforce diversity and negatively impact marginalized groups.

Potential solutions to these issues are to listen, learn and educate about health equity and antiracist work and to identify and address barriers to care for patients in each radiology practice.

“Acknowledging the discriminatory history of radiology and striving to improve diversity and health equity will ultimately work to improve patient outcomes,” the authors conclude.

Read the full article and accompanying invited commentary at

This article is also available for CME at Follow the RadioGraphics editor on Twitter @RadG_Editor.

Fig 5 Goldberg

Photograph of the National Medical Association Welcome Group, undated. (Image courtesy of Xavier University of Louisiana Archives and Special Collections, New Orleans, Louisiana.) ©RSNA 2023

Imaging Care for Transgender and Gender Diverse Patients: Best Practices and Recommendations

Although visibility and acceptance are improving for transgender and gender diverse (TGD) people, many still avoid seeking necessary medical care because of fears of discrimination or mistreatment.

In addition to encountering diagnostic errors caused by a lack of knowledge about TGD health, TGD patients experience disparities specific to radiology. Physician and staff ignorance is a barrier to effective care and unfairly transfers responsibility to the patient. One-third of TGD patients report having to educate imaging staff to receive appropriate care.

In an article published in RadioGraphics, Crysta B. Iv Kyrazis, MD, University of Michigan Health System, Ann Arbor, and colleagues, summarized the recommended best practices for patient care for TGD patients.

As a field, radiology has a range of opportunities for improving care delivery for TGD patients. This can be accomplished by creating a welcoming environment so patients feel safe during the imaging experience. Examples include using gender-neutral signage and all-gender single-user dressing rooms and bathrooms. Additionally, radiologists should be aware of reporting considerations for TGD patients, such as avoiding the use of gender in reports when not medically relevant and using precise, respectful language for findings related to gender-affirming hormone therapy and surgical procedures.

“Radiology departments have opportunities to optimize inclusive care for TGD patients at every stage in the imaging process. Radiology is important to patient care and we have the capacity to change TGD patient experiences and outcomes for the better,” the authors conclude.

Read the full article and accompanying invited commentary at

This article is also available for CME at Follow the RadioGraphics editor on Twitter @RadG_Editor.

Fig 3 Stein

Illustration shows penile inversion vaginoplasty, a feminizing surgical procedure in which the penile (purple) and scrotal (lavender) skin are used to construct the neovagina. The remaining scrotal skin is used for labiaplasty, and a sensate neoclitoris is created using the glans penis and penile neurovascular bundle. The prostate and seminal vesicles remain in place. ©RSNA 2023

Automated Inline Myocardial Segmentation of Joint T1 and T2 Mapping Using Deep Learning

Parametric mapping allows quantitative cardiac tissue characterization through calculations of local T1 and T2 relaxation times. Current practice involves acquiring T1 and T2 maps in separate acquisitions, which may require breath holds and prove difficult for some patients. These images are also often tedious and difficult for human experts to delineate.

In a new retrospective study in Radiology: Artificial Intelligence, James P. Howard, MA, MBBChir, PhD, Imperial College London, and colleagues used a joint T1 and T2 mapping sequence to acquire 4,240 free-breathing maps from 807 patients across two hospitals. Of those, 509 maps from 94 consecutive patients were assigned to a holdout testing set.

A convolutional neural network (CNN) was trained to segment the endocardial and epicardial contours using an edge probability estimation approach. After segmentation and mapping, the researchers compared network segmentation performance and segment-wise measurements on the testing set. The AI-derived measurements correlated closely with those of two human experts for native T1 maps, postcontrast T1 maps and T2 maps. For each measure, the interexpert correlation coefficient was within the range of the AI-expert agreements. With the use of simultaneously acquired T1 and T2 maps as the AI input, the segmentation failure rate dropped from 3.8% with T1 maps alone to just 0.8%.

“An AI solution using an edge probability estimation approach with a CNN allows automated quantitative tissue characterization using free-breathing T1 and T2 maps. Its performance is comparable with that of human experts,” the authors conclude.

Read the full article at and follow the Radiology: Artificial Intelligence editor on Twitter @Radiology_AI.