The author, a radiologist for one of the top Artificial Intelligence (AI) companies, Aidoc, discusses the challenges of implementing AI algorithms in radiology departments. The author uses the analogy of their past experiences repairing motorcycles to explain how deploying AI in healthcare settings often involves a collage of makeshift solutions reminiscent of duct tape, rather than professional, factory-made parts. Existing hospital systems, often built in the early 2000s or 1990s, are reticent to adapt to new technology, hindering innovation and making the integration of AI a difficult task.
AI researchers often gather extensive data, develop algorithms for specialized tasks (such as detecting which renal lesions in CT scans are likely to be renal cell carcinoma), and fine-tune these algorithms to a high degree of accuracy. However, the deployment of such solutions is often met with resistance. In many cases, this resistance stems from the outdated technology that constitutes the backbone of many medical systems which lack the capability to incorporate AI without risking system crashes or unacceptable latency.
To overcome these hurdles, many AI researchers have to resort to crafting bespoke solutions that enable the adaptation of their algorithm to the existing system. However, this process often leads to a less than optimal outcome, akin to using duct tape and cement to fix motorcycles.
The author points out the shift from predominantly single-algorithm, university-produced AI offerings, to a growing market of companies, like Aidoc, who develop and deploy a suite of AI algorithms across hospital systems globally. Aidoc, in particular, has had notable success in this area, with the most FDA-certified solutions in the industry.
Aidoc’s success is partially attributed to the development of a sophisticated AI delivery platform capable of deploying and supporting its algorithms across diverse medical settings globally. This platform can handle the vast amount of custom deployment challenges and avoids the ‘duct tape’ solutions of the past. Furthermore, it allows new algorithms to be easily integrated, which is crucial given that most of the currently commercial algorithms focus on prevalent issues and neglect less common, yet critical conditions.
The author concludes by emphasizing the importance of enabling the rapid growth of ‘low-prevalence’ algorithms through platforms like Aidoc’s, potentially akin to the medical industry’s ‘orphan drugs’. The continued evolution and broader deployment of such algorithms are critical for patient care and demonstrate the transformative potential of AI in healthcare.
