Research Challenges Limit Machine Learning Use in Medical Imaging

Though research on machine learning use in medical imaging has grown significantly in recent years, improvements in the clinical use of such data remain limited, according to a study published in npj Digital Medicine.

Machine learning (ML) is a promising but controversial tool for healthcare providers. Studies suggest heightened enthusiasm around the potential application of ML in clinical settings, but they also note that appropriate regulations must be implemented to ensure that it is effectively implemented. Recent studies have shown that biases within artificial intelligence (AI) algorithms can create health disparities.

The current study's authors found that at each step of the research process, potential challenges and biases can be introduced that limit the clinical use of ML in medical imaging. Issues can arise from the beginning, depending on how data for this research are collected, how datasets are created and distributed, and what biases may exist in the datasets themselves.

When the data are evaluated, further challenges present themselves in selecting meaningful evaluation targets, preventing improper evaluation procedures, choosing suitable metrics, and adopting other statistical best practices, the researchers pointed out.

At the publishing stage, certain incentives may impede the usability of the information presented. For example, authors may use elevated language and math to impress other academics, leading to a lack of clarity and omission of important details. Pressure to publish papers with “novel” methods and positive results can also lead to researchers using overly complex methods. These factors all reduce the reproducibility of a given study, which is key for determining whether the results are consistent and viable for further use.

To combat these challenges, researchers suggest raising awareness around data limitations, encouraging the use of established best practices for machine learning evaluations, and improving publication norms surrounding reporting and transparency.

Despite these concerns and concerns about the use of AI in medicine more broadly, there have been notable successes for ML recently.

One new study demonstrates that machine-learning models can help detect abdominal hernia surgery complications with high accuracy. Overall, these models predict hernia recurrence with 85 percent accuracy, surgical site occurrence with 72 percent accuracy, and 30-day hospital readmission with 84 percent accuracy. Along with improving patient outcomes, the research indicates that a 1 percent reduction in the rate of hernia recurrence, which these models could help facilitate, could save the US healthcare system $30 million.

Another new machine-learning algorithm helped clinicians flag high-risk colorectal cancer patients. The algorithm utilizes factors such as age, gender, and recent outpatient complete blood counts to determine which patients are at higher risk for developing colorectal cancer. Nurses can use the algorithm to schedule colonoscopies for these patients. Of the 68 percent of patients that scheduled a colonoscopy over the course of the study, 70 percent had a significant finding during the procedure.