(Research Use Only – Neuroradiology available, Cardio/Thoracic, Abdomen/Pelvic coming soon)
RadReport is a Machine Learning based Clinical Decision Support (CDS) tool that seamlessly introduces quality control and efficiency into radiology workflow. RadReport uses a proprietary Bayesian Expert System (BES) to arrive at a probabilistic Differential Diagnosis (DDx).
The Problem with Radiology Workflows today
Today, medical imaging still relies on the radiologist to extract useful information from image data and encapsulate that in the Report. Typically 100 words or less, this Report describes key image features observed by the radiologist (“Findings”) and a list of most likely diagnosis (“Impression”).
RadReport addresses several key needs of this existing radiology workflow.
- The current method of generating reports using Voice Recognition (VR) to create unstructured prose reports limits report quality and wastes at least 20% of a radiologist’s time. Current prose reports are highly variable, qualitative, and computationally barren.
- There are no checklists to remind radiologist of critical KFs to observe and there is no clinical decision support to help develop DDx.
- The radiologist not only spends time transcribing repetitive descriptions and diagnosis but also performing what were previously clerical tasks, including the correction of frequent VR transcription errors.
How it works
- Guides radiologist with validated image feature checklist
- Reduces observation errors while increasing speed
- Suggests most likely conditions, including rare diagnoses
- Reduces inference errors while increasing speed
- Drafts report and inserts into Reporting software
- Dramatically increases speed and report consistency
RadReport prompts the Radiologists with a validated set of image Key Feature (KF) inputs. It then recommends the most likely diagnoses and prepares a draft radiology report. The radiologist may modify the stated diagnoses and add any detailed descriptions specific to the case at hand.
RadReport integrates analytical capabilities into conventional radiology transcription systems, automatically populating report templates in real time.
Benefits of RadReport
RadReport helps the radiologist in the following ways
- It helps them more efficiently create consistent, higher quality reports by providing a checklist of carefully predefined KFs to feed the proprietary Bayesian Expert System (BES).
- The probabilistic Differential Diagnosis (DDx) generated serves as an additional check confirming the Radiologist’s own diagnosis, and reminding them of diagnoses they may not have thought about.
- These KFs and DDx are incorporated into the Findings and Impression sections of the report, using a natural language generator that produces concise, easily readable, highly structured and computationally useful reports.
Importantly and in contrast to many previous CDS tools, RadReport provides a very efficient, intuitive and friendly User Interface (UI) that actually saves radiologists’ time, a critical need in the setting of increasing image volumes and decreasing radiology reimbursement rates.
RadCAD (Coming Soon)
RadCAD is a Deep Learning based algorithm that extracts Key Features from radiologic scans.
The Problem we are tackling
Medical image interpretation techniques have not kept up with medical image generation technology.
Imaging in general and biomedical imaging in particular has two distinct elements – image production and image interpretation. Both are necessary for any practical application. While intimately related and mutually necessary, image production and interpretation are actually separate processes.
Over the last 30 years there has been an ‘explosion’ of new and increasingly useful medical image production technologies with CT, MRI, US, PET and SPECT now joining x-radiographs in exquisitely incorporating signals from most body parts into visually compelling images.
Contemporary image production creates huge amounts of incredibly sophisticated data, from which useful information remains to be extracted in order to make clinical diagnosis through image interpretation.
Unfortunately, image interpretation has changed little since Roentgen’s time. The medical image interpreter is still the traditional human observer, albeit more highly trained and knowledgeable than the observer of 1900.
While specialized human observers, such as radiologists, perform their interpretative tasks remarkably well, they, by necessity, use and are fundamentally limited by the human brain and visual system. The psychophysics of this system document that while human observers can tolerate poorer quality, noisy images remarkably well, we require image data with relatively high signal and contrast to noise, have a limited dynamic reporting range (approximately 7), and suffer from heuristic decision processes that are highly variable and error prone. Peer review data based on second readings indicates that at least 1 to 2 % of clinical reports do not include a critical finding or important diagnosis, resulting in at least 4 million erroneous or incomplete radiology reports per year in the US.
Furthermore radiologists currently use highly variable, non-discrete and qualitative prose, limiting their reports’ communicative and scientific value. To compound the problem, current radiology interpretative workflow is inefficient. The average time for a radiologist to view and report a standard radiographic study is 3 minutes, of which at least 20% is wasted in the mechanics of report generation. Such inefficiency has become economically crippling to clinical practices as reimbursement rates for medical imaging exams have been reduced by 30% since 2007. From both a supply and demand perspective, operative inefficiency has become the number one challenge to clinical radiology. A relatively fixed radiologist work force; a persistent annual increase in image volumes; and simultaneous reductions in per-exam reimbursements has led to greater and greater pressure on radiologist’s productivity.
The need for more efficient generation of better quality reports cannot be met without new interpretative technologies such as those offered by RadCAD.