Artificial Intelligence promises to revolutionize radiology. From early diagnostics to streamlined workflows and improved patient outcomes, AI holds immense promise. However, unlocking this potential hinges on one crucial factor: Responsible AI.

Building Trust in AI Models

The first challenge in the adoption of AI Radiology is ensuring that the AI output can assist the radiologist in the clinical workflow. The way the AI output is presented to the radiologist for consumption should improve her efficiency and reduce the time taken and any potential errors. Consuming the AI should not slow down the Radiologist by creating overhead. It should, in fact, be seamlessly integrated into the clinical reporting workflow. This should be addressed by using principles of Explainable AI. This ensures that the AI output can be used by the radiologist seamlessly as part of the clinical workflow by assisting the Radiologist and improving their workflows.

AI model outputs, across the industry, though may not be without false positives and negatives, are nonetheless improving day by day. The models today are better than what they were in the past and they will continue to improve in the future. Even with their current accuracies, using AI assistance in radiology reporting can add value to the reporting process and patient outcomes. For the same, the key is to first understand the performance of the AI model objectively. It is important to understand how the AI model performs across various ages, genders, subgroups, images from various modalities, etc. However, this visibility into the performance of the model across the hospital’s specific cohort is grossly missing.

The Benefits of Post-Deployment Performance Monitoring

Post-deployment performance monitoring offers a solution to these challenges and can unlock the full potential of AI in healthcare. Here are some key benefits:

Post-deployment Evaluation: Traditionally, the decision to use the model is based on a point-in-time, static evaluation done before deploying the model. Post-deployment monitoring introduces the ability to track the performance of the AI model after deployment.

Continuous Evaluation: The only way to measure the performance of the AI model post-deployment is to perform a costly offline validation exercise. Continuous evaluation enabled measuring the performance for every scan being reported using the AI model’s assistance without any additional effort.

Real-time Evaluation: Real-time evaluation enables monitoring the performance of the deployed AI models at any given time for all scans reported up to that time.

Ashutosh Pathak, CTO of DeepTek.ai