Greg Freiherr, Industry Consultant

Greg Freiherr has reported on developments in radiology since 1983. He runs the consulting service, The Freiherr Group.

Blog | Greg Freiherr, Industry Consultant | PACS | February 10, 2016

Watson, Come Here … Radiology Needs You

Watson, IBM, smart machines

Watson avatar courtesy of IBM

Fatigue. Bias. Trouble keeping up with peer-reviewed literature. These can spell trouble for radiologists. But machines are prone to none of them. Possibly best of all, artificially intelligent machines can fill in where needed … doing what has to be done, regardless of the circumstances.

IBM is counting on Watson Health to be one of those machines.

The company is building Watson Health, its artificial intelligence (AI) platform, to perform tasks that will benefit radiologists. Topping that list is summarizing the information in electronic medical records to provide the context that will help radiologists interpret images. A close second is an audit service that in one configuration looks for things that have been missed, in others expedites early interventions that can mitigate diseases that contribute to high mortality or cost. Cracking the top three on Watson’s list of possible applications is one that manages the radiologist’s worklist, providing radiological triage that moves cases up and down, depending on patient need. (A similar AI-enabled program, developed by vRAD, is nearing routine use. See “Will Artificial Intelligence Find a Home in PACS?”)

Ultimately these developers want to see Watson Health perform more sophisticated tasks, identifying cancers, circling suspicious masses and making the measurements that can help determine the effect of therapy or recurrence of disease.

The goal is to make the most of medical images and improve patient care, so as to help radiologists, not take their place, said Steven Tolle, an executive at IBM/Merge, the corporate conglomerate formed through the acquisition last fall of Merge Healthcare.

“Watson is not diagnosing the patient and certainly not replacing physicians,” said Tolle, chief strategy officer and president of iConnect Network Services.

While the search to image-enable the enterprise has helped many physicians by providing them with images that in the past had been only available to radiologists, not much has been done for radiologists. Watson’s EMR application would change that. Scheduled to be ready for use this year, it could find and summarize for radiologists information contained in the medical record, providing patient demographics, history and laboratory results. Watson Health might present these along with images from the patient’s prior exams.

IBM/Merge is not the first to work on a bridge from radiology to the EMR (see “Going Along to Get Along”). But Big Blue is the first to use AI in its construction.

Also scheduled for 2016 is a Watson Health “audit service,” Tolle said. This AI-enabled diagnostic workstation might help, for example, by prompting referring physicians to do follow-ups recommended by the radiologist.

“The radiologist might have read a chest CT, seen a nodule, and recommended follow-up in six months,” Tolle said. “We want to make sure that somebody schedules a follow-up exam of that patient.”

IBM plans to implement this audit service for high-priority diseases — top priorities are cardiovascular disease; cancers of the breast, skin, lung; and COPD— as well as diseases that impose high costs of care, such as diabetes and Alzheimer’s.

Watson Health will also provide technologies designed to make radiology more efficient and more effective. In the near future, Watson may juggle the order of cases on the radiologist worklist to give precedence to emergent needs. Brain bleeds would elbow their way to the top of the worklist, boosted there by a filter that recognizes patterns associated with intracranial hemorrhage.

To be of further help, Watson is being trained to provide access to peer-reviewed journals and medical texts relevant to the case being interpreted. Using pattern matching algorithms, similar to those used for facial recognition, Watson might look for images in a database that appear similar. These might serve as examples presented to the radiologist along with diagnostic “suggestions,” as Tolle calls them, generated by Watson.

“We want to provide options for the physician to consider,” he said.

Efficiency would be further improved, if Watson could learn to take over highly skilled but time-consuming tasks such as identifying and outlining masses, measuring them, and transposing the measurements into the radiology report. Initially Watson’s suggestions would be presented separately from the image being interpreted by the radiologist in another window or on a different monitor. The radiologist would always control what suggestions are accepted.

“Today the radiologist does this,” said Tolle, noting that in the future this might be delegated to Watson. “We are trying to streamline the radiologist’s workflow.”

Editor’s note: This is the second blog in a series of four by industry consultant Greg Freiherr on Working With Smart Machines. You can read the first blog, Will Artificial Intelligence Find a Home in PACS?, here.

Related Content

Kaplan-Meier curves for all-cause mortality in patients with versus without myocardial injury (Panel A) and in patients with versus without myocardial injury according to the presence or absence of major echocardiographic abnormalities (Panel B). *Includes wall motion abnormalities, global left ventricular dysfunction, diastolic dysfunction, right ventricular dysfunction and presence of pericardial effusion. Event rates are censored at 20 days from hospital admission. Images courtesy of Mount Sinai Health S

Kaplan-Meier curves for all-cause mortality in patients with versus without myocardial injury (Panel A) and in patients with versus without myocardial injury according to the presence or absence of major echocardiographic abnormalities (Panel B). *Includes wall motion abnormalities, global left ventricular dysfunction, diastolic dysfunction, right ventricular dysfunction and presence of pericardial effusion. Event rates are censored at 20 days from hospital admission. Images courtesy of Mount Sinai Health System

News | Ultrasound Imaging | October 28, 2020
October 28, 2020 — Card...
Ceiling-mounted X-ray system includes MyExam Companion intelligent user interface to guide technologist through exam workflow
News | X-Ray | October 26, 2020
October 26, 2020 — Siemens Healthineers has announced the Food and Drug Administration (FDA) clearance of the...
This illustration show the complexity of the data obtained from one single patient with moderate/severe traumatic brain injury. Different imaging approaches and techniques have their own unique sensitivity in assessing different aspects of neuroanatomy and neuropathology. What can be seen on images also changes with time since injury. Data from comprehensive clinical and functional assessments using a range of other tools is also important for evaluating patient outcome. Through data harmonization and large

This illustration show the complexity of the data obtained from one single patient with moderate/severe traumatic brain injury. Different imaging approaches and techniques have their own unique sensitivity in assessing different aspects of neuroanatomy and neuropathology. What can be seen on images also changes with time since injury. Data from comprehensive clinical and functional assessments using a range of other tools is also important for evaluating patient outcome. Through data harmonization and large-scale analyses of data combined across multiple research sites, the ENIGMA Brain Injury will develop and test methods and procedures for making sense of the complexity in this data. Images courtesy of Olsen et al., Brain Imaging and Behavior, 2020

News | Magnetic Resonance Imaging (MRI) | October 23, 2020
October 23, 2020 — Trau...
A, Initial conventional axial CT image shows no noticeable lung damage (within red box) in right upper lobe. B, Electron density spectral CT image obtained at same time as image in A shows lesions (within red box) in right upper lobe. C, Follow-up conventional axial chest CT image obtained 5 days after images in A and B confirm presence of lesions (within red box) in right upper lobe.

A, Initial conventional axial CT image shows no noticeable lung damage (within red box) in right upper lobe. B, Electron density spectral CT image obtained at same time as image in A shows lesions (within red box) in right upper lobe. C, Follow-up conventional axial chest CT image obtained 5 days after images in A and B confirm presence of lesions (within red box) in right upper lobe. Image courtesy of the American Roentgen Ray Society (ARRS), American Journal of Roentgenology (AJR)

News | Coronavirus (COVID-19) | October 22, 2020
October 22, 2020 — According to an open-...
The fMRI hyperscanning environment.

(A) The fMRI hyperscanning environment. The clinician (1) and patient (2) were positioned in two different 3T MRI scanners. An audio-video link enabled online communication between the two scanners (3), and video images were used to extract frame-by-frame facial expression metrics. During simultaneous acquisition of blood oxygen level–dependent (BOLD)–fMRI data, the clinician used a button box (4) to apply electroacupuncture (EA) treatment (real/sham, double-blind) to the patient (5) to alleviate evoked pressure pain to the leg (6; Hokanson cuff inflation). Pain and affect related to the treatment were rated after each trial. (B) Study overview. After an initial behavioral visit, each individual participated in a Clinical-Interaction (hyperscan preceded by a clinical intake) and No-Interaction condition (hyperscan without a preceding intake), in a counterbalanced order, with two different partners. (C) Experimental protocol. Each hyperscan was composed of 12 repeated trials (four verum EA, four sham EA, and four no treatment) in a pseudo-randomized order. After a resting period (far left), both participants were shown a visual cue to indicate whether the next pain stimulus would be treated (green frame) or not treated (red frame) by the clinician. These cues prompted clinicians prepare to either apply or not apply treatment while evoking corresponding anticipation for the patient. Following the anticipation cue, moderately painful pressure pain was applied to the patient’s left leg, while the clinician applied or did not apply treatment, respectively. After another resting period, participants rated pain (patients), vicarious pain (clinicians), and affect (both) using a visual analog scale (VAS).

News | Clinical Trials | October 22, 2020
October 22, 2020 — The potential impact of the patient-clinician relationship on a patient's response to treatment is
Lesion was originally reported as indeterminate enhancing mass, and outside report recommended biopsy. Classic features of benign hemangioma are shown. Error was attributed to faulty reasoning. A, Axial MR image obtained 5 minutes after contrast agent administration shows peripheral nodular discontinuous enhancement. B, Axial MR image obtained 10 minutes after contrast agent administration shows centripetal progression of enhancement (arrow). C, Axial fast imaging employing steady-state acquisition (FIESTA)

Lesion was originally reported as indeterminate enhancing mass, and outside report recommended biopsy. Classic features of benign hemangioma are shown. Error was attributed to faulty reasoning. A, Axial MR image obtained 5 minutes after contrast agent administration shows peripheral nodular discontinuous enhancement. B, Axial MR image obtained 10 minutes after contrast agent administration shows centripetal progression of enhancement (arrow). C, Axial fast imaging employing steady-state acquisition (FIESTA) MR image shows lesion is homogeneously hyperintense compared with liver parenchyma. Image courtesy of American Roentgen Ray Society (ARRS), American Journal of Roentgenology (AJR)

News | Magnetic Resonance Imaging (MRI) | October 21, 2020
October 21, 2020 — According to an artic...
According to an inquest, a man with a heart disorder and chest pain died two days after a doctor viewed the wrong scan and sent him home
News | Computed Tomography (CT) | October 21, 2020
October 21, 2020 — The BBC News
Flowchart of patient inclusion and exclusion.

Figure 1. Flowchart of patient inclusion and exclusion.

News | Coronavirus (COVID-19) | October 20, 2020
October 20, 2020 — A new multi-institutional study published in the journal ...
Rensselaer, First-Imaging, and GE Global researchers develop a deep neural network to perform nearly as well as more complex dual-energy CT imaging technology
News | Computed Tomography (CT) | October 20, 2020
October 20, 2020 — Bioimaging technologies are the eyes that allow doctors to see inside the body in order to diagnos
Lesion was originally reported as indeterminate enhancing mass, and outside report recommended biopsy. Classic features of benign hemangioma are shown. Error was attributed to faulty reasoning. A, Axial MR image obtained 5 minutes after contrast agent administration shows peripheral nodular discontinuous enhancement. B, Axial MR image obtained 10 minutes after contrast agent administration shows centripetal progression of enhancement (arrow). C, Axial fast imaging employing steady-state acquisition (FIESTA)

56-Year-Old Woman With Benign Hemangioma: Lesion was originally reported as indeterminate enhancing mass, and outside report recommended biopsy. Classic features of benign hemangioma are shown. Error was attributed to faulty reasoning. A, Axial MR image obtained 5 minutes after contrast agent administration shows peripheral nodular discontinuous enhancement. B, Axial MR image obtained 10 minutes after contrast agent administration shows centripetal progression of enhancement (arrow). C, Axial fast imaging employing steady-state acquisition (FIESTA) MR image shows lesion is homogeneously hyperintense compared with liver parenchyma.

News | Magnetic Resonance Imaging (MRI) | October 16, 2020
October 16, 2020 —