Bonding Radiologists with AI: How Radiologists will thrive
Shifting the narrative from Survival to the Thriving of Radiologists in the dawn of AI
Over the last decade, we have observed the fast proliferation of machine learning and data mining in a broad spectrum of industries. A branch of machine learning called Deep Learning has made significant inroads in the healthcare field of radiology.
Recently, Google DeepMind published an article in Nature where they conducted an extensive study on the detection of breast cancer from mammograms using Artificial Intelligence (AI). Specifically, DeepMind evaluated the diagnostic decisions made by radiologists and AI-Systems. Their results demonstrated that the AI-Systems were on par and in some cases outperformed the human radiologists. New York University had previously carried out a similar research study, which was also based on the identification of breast cancer from mammogram exams using deep learning. The outcome of the study illustrated that deep learning delivers comparable prediction results to human radiologists, and a hybrid deep learning enhances the performance of a human radiologist.
It is a standard and very common practice in machine learning research to compare the outcome of a research study with the status-quo, in order to properly evaluate the quality and scientific performance of an algorithm. Thus, it is justifiable for DeepMind and New York University to evaluate the quality of their deep learning breast cancer identification algorithms by directly comparing them with the performance of the status-quo — which in this case happens to be a human radiologist.
Like in other fields such as in Self-Driving Cars, Industry 4.0 etc, each time an AI solution is compared to humans for research or other reasons, public opinion inherently switches to questions such as — when is AI poised to replace humans in that field.
A similar narrative has dominated the conversation of AI in Radiology. Specifically, the narratives of many subsequent opinions (e.g., this piece) that referenced the aforementioned DeepMind article are echoing a gladiator fight between AI and radiologists, where only one must survive and there is little room for co-existence in the future. While such concerns are legitimate and have been arguable exacerbated by how the DeepMind’s Nature article seems to be specifically conveying a message that AI is better than human radiologists, there is a much brighter future for the co-existence between AI and Radiologists. Here is why.
Several online articles continuously point out how human radiologists are prone to errors, which may stem from fatigue or huge workloads.
However, from a machine learning perspective, AI-systems are also susceptible to errors. For instance, a lesion that can at times be immediately spotted by human radiologists can be missed by an AI-system due to the lack of diversity in training data. Thus, causing failures in a real-world production system. Based on this and other legal factors, final radiology diagnostic decisions should not be left to AI-Systems, no matter how good the scientific results of some studies appear to be. An AI-system should only act as an assistant to human radiologist, and a human radiologist should review, take responsibility and approve the final diagnostics result.
Succinctly, both human radiologists and AI-systems offer important complimentary aspects, which when combined can significantly enhance diagnostic accuracy and patient care quality. There is little need to pit one against the other, as this would lead to anxiety, reluctance and hesitance by many radiologists to quickly adopt AI in their radiology workflows. To pursue new breakthroughs in radiology, data scientists need radiologists expert knowledge to craft better AI models.
On the other hand, radiologists should not unwittingly limit themselves to their highly appreciated expert knowledge or the existing Computer Aided Diagnostic (CAD) software. Recent inferences and concerns that AI is poise to replace human radiologists is not true and simply an overreaction.
Radiologists need to forge a bond with AI and ignore such fears and concerns — as there are still surmountable number of early disease detection challenges in the brain, chest and breast that cannot be captured by human eyes.
This is an area where a bond between AI and radiologists can lead to new breakthroughs, clearly depicts how instrumental radiologists are required in the future, reduce human suffering and save lives. Radiologists will thrive in the future and drift towards full potential if both AI and radiologists co-exist and collaborate. Hence, the future should encourage radiologists to shape and govern AI policies, as well as to join force with AI to address challenging early disease detection problems. AI is there to foster the evolution and not a revolution in radiology. Who is better positioned to shape AI in radiology, other than the radiologists themselves.
Furthermore, to obtain a dramatic drop in health care cost, it’s insufficient to focus only on the deep learning algorithms that can detect and predict radiology diagnostic results. Instead, the entire radiology workflow has to be revisited and re-imagined from an AI and big data perspective. Such a comprehensive review would optimize the efficiency of the radiology workflow. In particular, big data and AI will optimize and automate the processes from medical scan generation through disease diagnostics by radiologists to diagnostics report delivery to the referring physicians.
For instance, Evercot AI, provides a cloud and on-premise Radiology AI-Platform called Greey Matter, that seamlessly injects and integrates AI in a hospital’s radiology workflow using big data technologies and AI. The AI-Platform provides cutting edge machine learning and deep learning solutions for the early detection of brain diseases such as Stroke, Alzheimer, Parkinson’s and Glioma.
Some other companies that focus on the use AI for radiology diagnostics include Aidoc, Zebras Medical, Enlitic and Viz.ai.
