AI Is Starting to Change Radiology, for Real
by Enhao Gong, PhD
The AI transformation is remarkable for its speed. For example in conference such as RSNA (Radiology Society in North America) and ISMRM (International Society of Magnetic Resonance in Medicine), AI applications was still a niche area before 2016, but is the top-1 hottest topic right now. We have passed beyond the point that people doubted the applicability and potentials of AI in radiology. Machine Learning and Deep Learning algorithms have demonstrated good performance that can supplement and verify the imaging tasks for radiologists. Although still in early stage, variable AI algorithms have already started to be tested in clinical environment. This year 2018 is also a landmark year for AI in radiology, since several AI+radiology products have received FDA’s nodding that they can practice in clinics for segmenting heart from MRI, make classification from CT, evaluate lung nodules, etc.
Beyond Classification – AI Improves Entire Radiology Workflow
Particularly, when talking about AI for radiology, most people think of an AI algorithm that can eventually “replace” radiologists and conduct diagnosis. When people talk about the possibility of “replacing” radiologist, it is important to recognize the complexity of radiology. Radiologists are imaging experts with more than 10+ years of medical training, who render diagnosis and clinical decision suggestion from interpreting medical images and correlating their findings from images with other exams and tests. Currently, most AI algorithms are often designed for ImageNet-ish image category classification tasks, to separate images from different categories such as with or without certain type of abnormality. It is clear that there are still huge gaps between predicting a classification label, which is well studied and achieved by AI algorithms, and rendering final diagnosis, which is much more complicated and requires both images and non-image information.
Beyond image classification, however, AI actually has great potentials to improve entire imaging/radiology workflow. AI algorithms and products can supplement existing radiology solutions, with more efficient and accurate imaging acquisition, reconstruction, analysis, detection, diagnosis, and prognosis, which goes way beyond predicting labels of normal and abnormal tissues.
Several research works, including the works we have done at Stanford and Subtle Medical, recently demonstrate that Deep Learning can replace conventional iterative algorithms for more accurate medical imaging reconstruction from 10x faster MRI scan or even 200x lower-dose PET scans. In addition, a lot clinical research works in neuroradiology; showed AI algorithms can not only categorize current disease, but also predict how the disease will progress in the future. From faster exams to accurate prognosis prediction, the entire radiology workflow can be improved using AI.
Different from other AI+radiology companies and startups focus on automating radiology diagnosis using AI, we at Subtle Medical chose a different route to integrate AI into clinical radiology. We provide hospitals and imaging centers software infrastructure for AI-empowered imaging workflow enabling hospitals to provide 4x faster and safer MRI and PET exams. We believe these AI tools will help hospitals and imaging centers to improve their productivity, quality of the service and patient satisfaction.
No Workflow Disruption from Disruptive AI Technology
It is no doubt that AI is a disruptive technology that will reform the radiology practice and workflow. However, the innovation in radiology practice, as well as in healthcare in general, should never disrupt clinicians’ operation. There is a saying that if the change of workflow requires radiologists to move from their chairs, it is not going to works. AI is supposed to free clinicians from repetitive tasks, not to add more tasks to them. Therefore it is fundamental for an AI product to seamlessly integrate into the entire workflow. For example, at Subtle Medical, all the products are designed in a way to (semi-) automatically function between scanners and PACS, staying almost silent and invisible to technicians and radiologists.
In general, we believe the best radiology tool should simplify, accelerate, and prioritize tasks for radiologists, making exams more accurate, efficient and personalized, and improve the productivity and satisfaction for both patients and clinicians. AI products have potentials to achieve all of these requirements.
Super-human Capability of AI
Several research works show state-of-the-art AI algorithms, such as Deep Learning, have great potentials in tasks human professionals are not good at as well. For example, in identifying diagnosis disease subtypes, similar to fine-grained image classification of identifying two similar dog breeds in ImageNet challenge, AI algorithms can do better than human and resolve inter-reader variability. In addition, for all the repetitive and quantitative measurements in radiology, such as contouring, segmentations, measuring brain thickness, and recording tumor size changes, AI algorithm can not only free clinician from these time-consuming mundane tasks but also get more accurate results. In addition, quantitative biomarkers and quantitative parameter mapping is a new trend in medical imaging as it offers better objective criteria for clinical decision, where Deep Learning is shown to improve for MRI.
Last but not least, new algorithms lead ways to correlate radiology with large scale genetic datasets, which is too complicated that human can never be good at but AI. All of these new advances from AI can contribute to personalized and precision medicine in future radiology.
Still a Long Way to Go
Although AI in radiology has achieved a lot progresses, there are still a long way to go until it is routinely used in clinics. In 2016, Dr. Geoffrey Hilton mentioned “it is quite obvious that we should stop training radiologists”. Later, several AI papers published on achieving “expert-level performance” in diagnosis add-up the concerns that radiologists should be worried about losing jobs. However nowadays we know there are definitely distances from “replacing” radiologists. As the developers of radiology AI, we also understand and appreciate more that radiologists are doing much more than image categorizing. Most AI/Computer-Assisted-Diagnosis products are still in their early stage for a closed-set of a few disease categories.
AI, for a foreseeable future, will not be completely generalized. No human is better than a small calculator for accurate arithmetic, but that is all that the calculator can do. Given the long-tail effects and the lack of datasets for rare disease, it is unlikely AI can do everything for radiologists. But, as put by Dr. Langlotz from Stanford, “radiologists who use AI will replace radiologists who don’t.” AI in radiology, probably similar to AI in other popular areas such as manufacturing and autonomous driving, will keep improve its capability to change the entire industry. As quoted in a recent Forbes article on robotics: AI will replace tasks, not jobs.