 (AI) adoption in radiology has, surprisingly, taken longer than many of us had hoped and anticipated. Many in the industry have learned with experience that this journey is not a sprint, but a rather a marathon with a relay-race component, because it requires a broad ecosystem of stakeholders committed to teamwork and collaboration. As a result, the industry’s perspective on AI in radiology is continuously evolving and maturing.

Five years ago, people were focused on how AI could address and relieve physician burnout, by helping the radiologist who was facing an ever-expanding workload. But, despite intense levels of burnout, increasing case volume and complexity and staffing shortages, AI adoption has not accelerated as originally expected. The American College of Radiology (ACR) Data Science Institute (DSI) conducted a study

 in spring 2020 and found about one-third of survey respondents reported using AI in their practice. Of those respondents who were not using the technology, only 20% planned to purchase it.

When ACR DSI dug a little deeper into why radiologists were not adopting AI, they found that 80% of survey respondents not using AI in their practices said that they “see no benefit” to using the technology.

 I believe this illustrates the crux of the matter: Is AI valuable to radiologists?

Transitioning from Entrepreneurial to Ecosystem Approaches

Early on, AI in radiology was driven in a very entrepreneurial way—similar to an “if you build it, they will come” model. Many developers built innovative solutions quickly, without necessarily assessing market needs or budget availability.

However, healthcare leaders will put any new innovation to the test, asking: Will it improve quality of care and outcomes? Will it reduce costs or make money? Will it make work easier, faster or better in some way? If it’s not doing any of those things, then the answer will likely be “no, thank you.”

So, despite the fact that many early AI algorithms were innovative and impressive, they may not have targeted common pain points well enough to succeed. I think the industry has also learned from this early, exciting innovation phase that developing and adopting AI is not a standalone process. Technologies must be integrated with existing platforms and workflows, as well as demonstrate value.

Defining value will require thinking deeply about the radiologist’s day-to-day experience and workflow. When I think about my own experience as a practicing radiologist, I can think of a couple of specific areas where AI could bring value to me. First, I look at a recurring task— oncology follow-ups—that show up consistently on my worklist. While not a difficult task, it is repetitive and time-consuming to do it well. Could AI help automate this task on my list? Another way I would find value in AI is in helping identify 

, especially if this capability is just part of my workflow and doesn’t require me to open yet another solution. And from my perspective, AI has real potential value to help with time-consuming, costly and important non-clinical tasks, such as workflow, quality assessment, patient tracking and follow-up, data summarization (especially unstructured data), and appropriate use. Those are a few of the places where I’d love AI to step in, working well with all of my existing tools, and help me. 

Defining the areas where AI can bring value to the radiologist is a critical first step. But when the definition of AI value flows this direction—from clinician to the developer—instead of the other way around as it did in the entrepreneurial model, we also need to re-think how we’re delivering AI.

Orchestration: An Evolving Approach to AI in Radiology

Enter AI orchestration, which is one way 

IBM Watson Health is evolving our approach to AI in imaging

. AI orchestration is a new way to help imaging organizations to experience the benefits of AI with the goal of having the AI applications working seamlessly together. This approach simplifies the end-to-end process of finding, procuring and implementing diagnostic imaging AI-enabled solutions. 

Because each AI application is developed in a unique way with a specific purpose, it can be challenging for organizations to review and assess each one, and then to deploy them in a way that’s beneficial to users. Orchestration offers a single platform that makes it easier to deploy a curated selection of AI applications, to consume their results, and to manage new innovations over time. This approach helps integrate the right toolset—both in the physician’s workspace as well as the IT environment.

, IBM Watson Health officially launched our IBM Imaging AI Orchestrator, a cloud-based imaging AI service that provides scalable access to regulatory-cleared AI applications. It is designed to help organizations manage imaging worklists by indicating status of AI processing and returning consolidated results from AI applications directly into the PACS. Physicians can use it to control their access to AI insights within the reading process, while IT can outsource the management of multiple applications to IBM.

Ultimately, I believe radiology will be better off when AI development and adoption engages all stakeholders—clinicians, developers, executive leaders, administrators, academia, organized medicine and other experts across the industry. If we think about the problems we need to solve together, we can bring greater value. The industry is starting to get there, and I expect great things yet to come.

Articles

RSNA

Technologies must be integrated with existing platforms and workflows, as well as demonstrate value.

Changing and Maturing Perspectives on AI 

Low-field magnetic resonance imaging (MRI) offers comparable performance to routine computerized tomography (CT) and MRI for following hydrocephalus. This is according to initial results from an ongoing trial presented at the 

Radiological Society of North America (RSNA) 2021 Annual Meeting

, which also found that deep-learning based lateral ventricle volume estimates from low-field MRI are accurate.

“Patients with hydrocephalus require repeated neuroimaging to monitor ventricular volumes and avoid complications,” the authors wrote. “Low-field MRI may offer a cost-effective, radiation-free alternative to monitor ventricular volume changes in patients with hydrocephalus.”

The results were presented by Thomas Campbell Arnold, DPhil, department of bioengineering at the University of Pennsylvania in Philadelphia.

Standard of care imaging for patients with hydrocephalus includes routine CT or MRI. However, CT uses ionizing radiation and provides less soft-tissue contrast and MRI is more expensive, less widely available and is subject to imaging artifacts from commonly used programmable ventricular shunts. Further, strong magnetic fields change programmable shunt settings.

“We hypothesized that low-field imaging would provide a comparable and more convenient point-of-care assessment for alterations in ventricle size, less shunt-related artifact and less need for programmable shunt checks,” the authors wrote.

In this ongoing study, the researchers assessed a recent US FDA cleared, portable, low-field strength (64mT) MRI scanner (Hyperfine, Guilford, Connecticut) that was used to monitor 22 patients with known or suspected hydrocephalus. Same-day low-field MRI and standard of care imaging, with CT or MRI, were collected for each patient. A neuroradiologist assessed ventricles as abnormally increased, decreased or within expected limits. Additionally, semi-automated ventricular volume segmentations were generated for clinical MRI scans and compared with deep-learning based volumes that were provided automatically by the low-field device.

Radiological assessments were comparable between low-field MRI and standard of care imaging. One patient presented with a shunt related hemorrhage on both imaging modalities. A strong correlation (R=0.97, p<0.001) was found between lateral ventricle volume measurements on low-field and clinical MRI.

Settings for programmable shunts, which were present in 10 patients, were recorded before and after low-field imaging. Shunt settings were altered in nine of these cases.

“Programmable shunts must still be checked by the clinical care team after low-field imaging,” the authors wrote.

Early study results suggest that low-field MRI may offer a cost-effective, radiation-free alternative to monitor ventricular volume changes in patients with hydrocephalus. 

Low-field MRI Matches Standard-Of-Care Imaging for Hydrocephalus

 describes a community-based outreach intervention that was effective at identifying and addressing barriers to lung cancer screening, such as cost, lower literacy levels and fear of diagnosis.

“Radiologists can partner with patients and key community stakeholders in the development of effective lung cancer screening outreach interventions aimed to increase equitable lung cancer screening participation among low socioeconomic status communities,” the authors wrote.

The results were presented by Efren Jesus Flores, M.D., department of radiology at Massachusetts General Hospital, in Boston.

Community-based lung cancer screening outreach can help individuals from low socioeconomic status communities, who experience worse lung cancer outcomes, by addressing barriers to digital health and promoting trust. In this study, the researchers aimed to develop a tailored digital health outreach intervention to promote lung cancer screening and understand related social needs among low socioeconomic status communities.

Focus groups were conducted at an urban academic medical center with affiliated community health centers with 12 representative primary care physicians, eight advocates, seven lung cancer screening patients and eight non- lung cancer screening patients. Three tailored lung cancer screening outreach video conditions were developed with messaging by a radiologist, patient and patient/radiologist together. 

A national sample of 315 current smokers with Medicaid who were eligible for lung cancer screening were randomized to watch one of the three video messages to determine satisfaction, preferred source, intent to screen and attitudes about lung cancer.

The focus groups cited transportation and fear about lung cancer as major barriers to screening. Patients reported confusion about screening eligibility criteria, the screening process and insurance coverage. The outreach video was modified to emphasize the treatability of early lung cancer, adopted clear language for eligibility and insurance coverage, and displayed a patient undergoing lung cancer screening and receiving results. 

Of the 315 participants, 67% reported satisfaction with the video, with those watching the video messaging from the patient reporting the greatest intent to speak to a physician about lung cancer screening. Participants reported limited social needs.

“Greater social need was associated with significantly lower health literacy, smoking-related stigma, lower confidence in arranging transport to lung cancer screening, greater cost concerns about screening, perceived severity of lung cancer and greater worry about developing lung cancer,” the authors wrote.

Cost, lower literacy levels and fear of lung cancer diagnosis were highlighted as concerns with lung cancer screening. 

Community-Based Intervention May Increase Lung Cancer Screening 

 suggest that diffusion-weighted imaging was a sensitive and effective screening method for breast cancer detection in asymptomatic women with dense breasts. 

“Mammography has low sensitivity in dense breasts and needs to be supplemented with another screening method for women with dense breasts,” the authors wrote. “Diffusion-weighted imaging may be useful as a safe and effective screening tool for the supplementation of mammography in women with dense breast tissues.”

The results were presented by Hee Jeong Kim, M.D., breast cancer center at Asan Medical Center in Seoul, Korea.

In the study, the researchers assessed the diagnostic performance of diffusion-weighted imaging for breast cancer detection in asymptomatic women with dense breast tissue on mammography. The researchers also compared the diagnostic performance of mammography, dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and diffusion-weighted imaging.

From August 2018 to January 2019, the researchers retrospectively reviewed 297 consecutive asymptomatic women, mean age of 51 years, with dense breast tissue on mammography who underwent breast MRI including diffusion-weighted imaging and DCE protocols and mammography. Three readers independently reviewed the lesion visibility and level of suspicion on each imaging modality. Histopathologic features of the tumors on surgical specimen served as the standard of reference.

A total of 233 breast cancers were identified in 230 women, benign lesions were identified in 11 women and normal breasts were identified in 56 women. Of the 233 cancers, 15% presented as ductal carcinoma in situ, 12% as microinvasive ductal carcinoma and 73% as invasive carcinoma. The overall sensitivity was 61.4% in mammography, 79.4% in diffusion-weighted imaging, 86.3% in DCE-MRI and 91.8% in the combination of mammography and diffusion-weighted imaging. The sensitivities for noninvasive cancer in mammography, diffusion-weighted imaging, DCE-MRI, and the combination of mammography and diffusion-weighted imaging were 73%, 68.3%, 74.6% and 92.1%, while the sensitivities for invasive cancer smaller than 20 mm were 51.9%, 81.5%, 88.9% and 89.6%, and the sensitivities for invasive cancer larger than 20 mm were 77.1%, 91.4%, 97.1% and 100%, respectively. 

When diagnostic performance was assessed by using areas under the receiver operating characteristic curves, diffusion-weighted imaging was superior to mammography (

 < .001) and comparable to DCE-MRI when combined with mammography (

Diffusion-weighted imaging may be useful as a safe and effective screening tool to supplement mammography in women with dense breast tissues. 

Diffusion-Weighted Imaging Helps Detect Breast Cancer in Dense Breasts 

 found that racial ethnic minorities, and patients with low socioeconomic status, lower access to broadband internet or limited English proficiency were less likely to use a patient-portal self-scheduling pathway for screening mammograms.

“Understanding the factors associated with patient use of technology and patient portals for access to care can guide interventions to bridge the existing digital divide and promote health equity in radiology,” the authors wrote.

The results were presented by Patricia Balthazar, M.D., assistant professor of abdominal radiology and imaging informatics at Emory University in Atlanta, Georgia.

While patient portals can increase patient access and engagement, they can create potential barriers to care. In this study, the researchers aimed to identify potential sociodemographic disparities in the use of an online patient portal to self-schedule screening mammograms compared with traditional scheduling pathway, such as a scheduler phone call and referral system. The researchers hypothesized that patients with limited English proficiency or low socioeconomic status, along with racial/ethnic minorities would be less likely to use the online self-scheduling pathway.

This retrospective cohort study included 46,268 female patients undergoing screening mammograms from Jan. 1, 2019, to Dec. 31, 2019, at an urban quaternary-care academic medical center. The following data were extracted: patient scheduling pathway, age, language, race, health insurance provider and zip code, with the latter linked to Census data to extract the following demographic information: internet access, median household income and education level.

Of the patients, 0.7% used the online self-scheduling pathway. Those using the self-scheduling pathway had higher odds of being younger (OR for age in year: 0.94; 95% CI: 0.93-0.96), English-speakers (95% CI: 21.6, 3.0-156.5), White (95% CI: 1.7, 1.2-2.5), have private insurance (95% CI: 1.5, 1.0-2.1) and to live in zip code areas with higher percentage of access to broadband internet (95% CI: 1.2, 1.1-1.3).

Unlike racial ethnic minorities, patients with low socioeconomic status, lower access to broadband internet and limited English proficiency, “patients living in zip code areas with higher median household income, education level or with any type of internet access were not significantly associated with the online self-scheduling pathway,” the authors wrote.

This study discloses the characteristics of patients using self-scheduled online patient portal screening mammography. 

Online Patient Portals for Screening Mammograms May Widen Health Disparities 

68Ga prostate specific membrane antigen (PSMA) positron emission tomography (PET) has high utility in the detection of suspicious lesions in biochemically recurrent prostate cancer. This is according to study results presented at the 

“PSMA PET is a more robust modality for biochemical recurrence with the potential of identifying small lesions which would otherwise be dismissed on MRI alone,” the authors wrote. “We recommend the use of 68Ga PSMA PET/MR for biochemical recurrent prostate cancer with potential to be the new standard of care for this indication.”

The results were presented by Juana Martinez, M.D., fellow in radiology at Weill Cornell Medical College in New York. 

PSMA PET combined with magnetic resonance imaging (MRI) is a promising technique in the evaluation of patients with prostate cancer. In this study, the researchers compared PSMA PET with concomitant multiparametric MRI in detecting lesions suspicious for biochemical prostate cancer recurrence. 

A review was performed of all patients with prostate cancer who prospectively enrolled in a clinical trial of 68Ga PSMA PET/MRI at New York Presbyterian Hospital - Weill Cornell campus from April 2018 to January 2021. Nuclear medicine physicians read the PSMA PET portion of the study and a radiologist with fellowship training in cancer imaging read the MRI of the skull base to thigh, independent of each other. Patient demographics, the indication for PSMA PET/MRI and serum prostate specific antigen (PSA) levels closest to the PSMA PET/MRI were extracted from patient charts.

Of the 165 patients who had a PSMA PET/MRI, 109 where imaged for biochemical recurrence. PSMA PET was able to detect more lesions suspicious for recurrence than MRI at all PSA levels. When assessing only patients who had PSMA PET/MRI for biochemical recurrent disease, PSMA PET fared superiorly at all PSA levels in comparison with MRI. 

“When compared with MRI, PSMA PET is highly robust in identifying lesions which are otherwise deemed unremarkable,” the authors wrote.

PSMA PET is highly robust in identifying prostate cancer lesions which are otherwise deemed unremarkable. 

Prostate-Specific Membrane Antigen PET Bests MRI in Biochemical Recurrent Disease 

The concurrent use of an artificial intelligence (AI) tool with digital breast tomosynthesis could improve the diagnostic performance of radiologists in the detection of breast cancer. This is according to a study presented at the 

 that also found that the addition of AI improved the efficiency of radiologists’ workflow.

“This AI system can be integrated into screening practice to improve both efficiency and accuracy of digital breast tomosynthesis,” the authors wrote.

The results were presented by Pierre Fillard, Ph.D., the founder and chief scientific officer of Therapixel SA in Nice, France.

The retrospective study aimed to determine the benefits that an AI tool could bring to the time and accuracy of digital breast tomosynthesis interpretation. A total of 22 breast specialist radiologists were asked to interpret 240 digital breast tomosynthesis examinations with and without the help of AI. The dataset included 114 biopsy-proven cancers, of which 36 were classified as false negative. Performances were measured in terms of area under the receiver operating characteristic curve (AUC), sensitivity, specificity and reading time.

The average AUC across readers was 0.79 without AI and 0.83 with AI, for an average improvement of 0.04 (95% CI: 0.02 to 0.06, 

 < .001). Average sensitivity improved from 80% without AI to 82% when using AI (2%; 95% CI: -0.4% to 0.4%), while specificity increased from 56% to 61% (5%; 95% CI: 1.5% to 8.7%). Average reading time was 74.7 seconds without AI and 70.9 seconds with AI (average difference: -3.8; 95% CI: -8.4 to 0.8). The reading time reduction varied in accordance with the likelihood of malignancy assigned by the AI, ranging from -15.6% for cases assigned with a low likelihood of malignancy to -0.5% for the most suspicious cases.

When applying these results to a screening population, the observed reading time was on average 76.2 seconds without the help of AI, which the authors noted “lies in the lower bound of the reading time reported for real-word practice.”

“Nonetheless, an average reduction of 8% was observed in assisted readers, which indicates that a regular batch reading (i.e., a sequential interpretation of screening mammograms without interruption) could include 13% more cases (about seven cases per hour) thus improving the efficiency of the screening program and reducing the time-to-diagnosis for patients,” the authors wrote.

An artificial intelligence for digital breast tomosynthesis enhances radiologists’ performance and efficiency. 

AI for Digital Breast Tomosynthesis Enhances Diagnostic Performance 

 suggest that magnetic resonance imaging (MRI) alone is a successful method for the follow-up in women who undergo breast-conserving treatment and ensures early detection of recurrent cancer.

“Women in the follow-up after breast cancer who have access to breast MRI may safely forgo additional mammographic or digital breast tomosynthesis follow-up,” the authors wrote.

The results were presented by Vanessa Raaff, Ph.D., of the department of diagnostic and interventional radiology, RWTH Aachen University Hospital in Germany.

For follow-up after breast conserving treatment, serial two-view mammography is recommended, possibly accompanied by targeted ultrasound. As breast MRI has been shown to be useful to improve breast cancer detection of recurrence in patients after breast conserving treatment, the researchers investigated the usefulness of MRI as a standalone imaging method for this purpose.

Between January 2011 and March 2021, a total of 426 women who had undergone MRI alone for routine follow-up after breast conserving surgery were included. The women started MRI screening at a median of two years after treatment and underwent MRI screening annually. Validation of the last negative MRI study was achieved by tomosynthesis plus clinical examination in 137 women and/or by clinical follow-up for at least two years in 289 women.

The women were followed by 1,553 MRI studies, covering a total 1,727 women years. Across the entire follow-up period, MRI was positive in 108/1,553 examinations (6.9%) and in 85/426 women (19.9%). Recurrent breast cancer was confirmed in 64/108 MRI studies and in 55/85 women, for a positive predictive value of 59.3% per positive MRI study and 64.7% per patient. 

The cancer-detection-rate was 37 per 1000 women years. Recurrent cancers were detected on average 8.9 years after breast conserving treatment. Of the 64 recurrent cancers, 17.2% were ductal carcinoma in situ and 82.8% were invasive. The mean size of invasive recurrent cancer was 11.4 mm, most frequently staged as T1b, followed by T1c.

“None of the women received a diagnosis of breast cancer in between MRI screening rounds, or within the two-year follow up after the respective last negative MRI, or upon digital breast tomosynthesis after the last study MRI,” the authors wrote.

Following breast cancer treatment, women who have access to breast MRI may safely forgo additional mammographic or digital breast tomosynthesis follow-up. 

MRI Alone Beneficial for Surveillance After Breast-Conserving Treatment

 describes a deep learning model that was developed to successfully detect intracranial hemorrhage in computed tomography (CT) scans without medical annotation.

“This method is faster because it removes the need for medical annotation, which can be added later on image errors to improve the model,” the authors wrote. “By detecting 96.6% of the normal exams, we can speed up doctor’s work, providing more time for the analysis of the slices with bleeding, increasing the precision of the diagnosis.”

Intracranial hemorrhage is a life-threatening condition that is diagnosed via CT scans. Deep learning algorithms developed to aid in the detection of bleeding within the intracranial cavity require a physician to annotate exams per slice. In this work, the researchers annotated a dataset of head CT exams using natural language processing (NLP) and applied an architecture using convolutional and recurrent neural networks to classify each exam. 

A total of 8432 head CT exams collected from six hospitals and five specialized outpatient clinics were assembled and the axial volumetric unenhanced series of scans were selected. The labels were extracted from the written reports using a previously trained NLP model. Almost 40% of exams had an intracranial hemorrhage, which were not separated by bleeding subtype. Each exam was processed using brain, blood and bone windows, resliced to 25 images. A MobileNet network pre-trained with ImageNet weights was combined with a Long short-term memory (LSTM) with categorical cross-entropy as loss function. The model’s output was as follows: hemorrhage, no finding and other pathologies. 

The model improved the classification without the need for a per slice annotation. On the validation set, an area under the curve (AUC) of 0.96, 0.87 for recall and 0.97 for precision was achieved. While on an external set, validated by a doctor, the AUC was 0.91, 0.85 and 0.86, respectively. 

The researchers also highlighted slices where the hemorrhage appears to aid diagnosis by applying Gradient Class Activation Map (GradCAM) technique. 

“Through the GradCAM method, we selected the best slices of each exam, and doctors were able to evaluate the results and indicate with more precision where a hemorrhage was found,” the authors wrote. “The identification of the most relevant slices, using the GradCAM technique, alongside clinical information, may help doctors better understand each patient case, accelerating and improving the diagnosis.”

A deep learning model was developed to detect intracranial hemorrhage in computed tomography scans without medical annotation.

Algorithm Can Detect Intracranial Hemorrhage in CT Without Medical Annotation 

In patients with esophageal squamous cell carcinoma (SCC), multiparametric magnetic resonance (MR) features provide useful diagnostic information of pathologic complete response after neoadjuvant therapy. This is according to a study presented at the 

“Multiparametic MR features provide useful diagnostic information of pathologic complete response among patients with esophageal SCC who undergo neoadjuvant therapy,” the authors wrote.

The results were presented by Jinrong Qu, M.D., of the department of radiology, affiliated cancer hospital of Zhengzhou University and Henan Cancer Hospital in Zhengzhou, China.

 earlier this year found that MRI has high sensitivity but low specificity for the detection of esophageal SCC and adenocarcinoma, showing promise for determining neoadjuvant therapy response. Meanwhile, a 

 of 54 patients with locally advanced esophageal SCC found that integrating parameters of diffusion-weighted MR imaging (DWI) with other imaging methods can better identify pathological response of primary tumor to neoadjuvant chemoradiotherapy.

The latest study presented by Dr. Qu included 36 patients with esophageal SCC at stage cT2-T4a/N0-N3/M0 who were retrospectively reviewed. MRI was performed before and after neoadjuvant therapy, followed by surgical resection. The lesions were analyzed for the following five MR features: changes of size, signal on DWI, apparent diffusion coefficient (ADC) map, T2-weighted turbo spin echo (TSE) BLADE and degree of enhancement. Response was assessed by using tumor regression grade (TRG).

The five MR features after neoadjuvant therapy showed strong positive correlations with TRG 0 (r = 0.7432; 

<.001). The 10 cases with TRG 0 showed either complete disappearance of the mass or decreased mass with minimal residual thickening of the esophageal wall. These TRG 0 cases also showed iso- or decreased-intensity on DWI, iso- or increased-intensity on ADC map, iso- or decreased-intensity on T2-weighted TSE BLADE, or homogeneous enhancement on high-resolution delayed StarVIBE. The five MR features were different from all the TRG 2-3 in 20 cases. 

ADC map had the highest performance in distinguishing TRG 0 from non-TRG 0 with the sensitivity of 96.1%, specificity of 76.5%, negative predictive value of 74.7% and positive predictive value of 91.6%

“MR assessment of changes in tumor size, signal on DWI, ADC map, T2-weighted TSE BLADE, and enhancement pattern may distinguish between biopsy proven TRG 0 and non-TRG 0 in patients with esophageal SCC,” the authors wrote.

Multiparametric MRI can detect pathologic complete response after neoadjuvant therapy in patients with esophageal squamous cell carcinoma.

Latest Conference Articles