The following article details the author’s experience of working at Aidoc, a leading medical AI company, despite lacking a detailed understanding of software engineering, data security, and AI, drawing parallels between his experience repairing old motorcycles and developing and deploying AI algorithms in medical settings.
The author introduces the topic by confessing his lack of comprehensive knowledge of the technical aspects of AI. He contextualizes his profession as a radiologist and how it provides a unique perspective in understanding AI alongside his work. The author focuses on the final step of algorithm development – releasing it into the wild, replete with its own challenges.
Developing an AI algorithm for medical purposes involves extensive data collection, algorithm design, recursive training, and algorithm testing. However, deploying them in a real-world context, let’s say, in a radiology department’s Picture Archiving and Communication System (PACS), often reveals incompatibilities. Challenges include securing permission to tweak the PACs servers and software, potentially encountering outdated technology, or lacking a suitable hardware connection for AI integration.
This process often results in a custom-built solution that enables the AI to function without disrupting the existing system – a process likened to the author’s experience fixing old motorcycles with makeshift parts. Despite the improvisation, such solutions remain suboptimal for a critical system like PACS or Electronic Health Record (EHR).
AI development has since evolved, with companies like Aidoc successfully deploying sets of FDA-approved algorithms on varied hospital systems worldwide. The experienced team at Aidoc has evolved their initial, makeshift approach to algorithm deployment into a sophisticated platform that can be rapidly and securely installed and supported in almost any setting. Such an innovative platform not only harmonizes well with large, complex hospital systems but also facilitates the seamless integration of new algorithms into the system.
Despite the success in diagnosing common medical anomalies using commercial radiology algorithms, it remains a challenge to utilize them for rare conditions. Algorithms for such conditions are often developed by academic medical centers and are seldom readily commercialized. The author concludes that these low-prevalence algorithms, analogous to orphan drugs, are universally important, and a supportive deployment platform could help improve them by accessing more case studies.
