For a while now, artificial intelligence (AI) has been touted as the replacement for radiologists. Geoffrey Hinton, the godfather of AI and a former scientist at Google Brain, said in 2016 that “It’s completely obvious that within five years, deep learning is going to do better than radiologists.” Deep learning is a branch of machine learning that eliminates some of data pre-processing to run an AI model.
A report published in the Lancet Oncology Journal on August 1 revealed that using AI in breast cancer screenings could halve workload of radiologists. The study followed more than 80,000 women from Sweden with the average age of 54, to directly compare AI screenings with standard tests. While AI-supported screenings detected 244 women (28%) with cancer, standard screenings found 203 women (25%) who had cancer. The study also showed that the use of AI did not generate more false positives than the standard tests.
A joint report, released in July, authored by researchers from MIT and Harvard Medical School, revealed that doctors actually did worse at diagnosis when working with AI. The authors noted when the AI system predicted a certain diagnosis, doctors became more certain about the opposite prediction, which was the wrong one.
Building Trust
For instance, when the AI system reported that a patient had ‘X%’ of a disease, doctors altered their own diagnosis in a way that reduced positive results, indicating a lack of trust. Doctors even performed worse when assisted by AI, and they were second-best while performing alone compared to an AI-alone diagnosis.
(For top technology news of the day, subscribe to our tech newsletter Today’s Cache)
While AI’s diagnoses are reliable, they are also prone to misjudgements.
“So, we use AI for speed and ease but we monitor its results before releasing the conclusion to the patient,” Dr. Pradeep Srinivasan, head of radiology at Fortis, Bannerghatta in Bangalore said. “AI is a double-edged sword and we have to use it with great safety protocols and it should be monitored always.”
But medical professionals are also gradually bonding with AI as “AI’s consistent performance in identifying critical markers and its ability to process a colossal volume of information” are fostering confidence in its capabilities, Dr. Shivakumar Swamy. S, head and director of radiology at HCG Cancer Hospital, Bangalore said.
AI is often used to refine workflow by cutting down time taken by making scans clearer. This facility has been extremely useful in avoiding unnecessary repeat scans, Dr. Prakashini K, Professor and head, department of radio diagnosis at Kasturba Medical College and Hospital said.
“Precise image module provides improved diagnostic quality images with significantly low patient radiation dose, making a positive impact on the delivery of value-added healthcare services,” she added, noting that AI is also used in live needle tracking during CT-guided biopsies, thereby boosting the confidence of the radiologist and the efficiency of the procedure.
(For top technology news of the day, subscribe to our tech newsletter Today’s Cache)
‘A second reader’
But data sets on specific diseases is a critical factor to know how effective an AI model is.
“Performance of AI in general depends on the amount of data sets of particular disease that has been used for making the algorithm and how well it is found to be representative of varied presentations of a particular disease,” she said. AI is still considered good enough to be a ‘second reader’ for helping in segmentation and identification of pathology to some extent. It requires more research and technical developments to become a reality.
The MIT and Harvard study concluded that when doctors were shown the images plus case history, their diagnoses were immediately better. On the other hand, AI systems had no knowledge of a patient’s medical history as it wasn’t trained on this data.
“Our trust in AI is not absolute. While it excels in pattern recognition and quantitative analysis, the nuanced art of understanding a patient’s unique medical history and context requires finesse of human intervention. AI does offer tremendous support, but our responsibility is to exercise clinical judgment and make decisions considering the broader spectrum of patient care. The future envisions a dynamic partnership between radiologists and AI, where each reinforces the other’s strengths,” Dr. Swamy added.
