Artificial intelligence is rapidly shaping the future of healthcare, particularly in radiology, where deep learning models uncover new possibilities in medical image analysis. However, taking an AI concept from ideation to real-world application is a challenging journey that requires navigating technical, regulatory, and clinical landscapes. In this blog, we draw insights from a recent webinar featuring Mehmet Yiğitsoy, Principal AI Engineer at deepc, and Javier García López, CTO and co-founder of Sycai Medical, to explore the key steps, challenges, and lessons learned in building impactful AI solutions for healthcare.
Javier’s journey with Sycai Medical began by identifying a critical need in healthcare: early detection of pancreatic cancer, a disease with a life expectancy of less than five months upon diagnosis. Through over 200 interviews with radiologists, gastroenterologists, and oncologists, his team uncovered a major challenge—managing pancreatic cystic lesions. Radiologists reviewed cases, manually measured lesions, and documented changes. This was a clear opportunity for AI to provide value by automating these processes and reducing workload.
Tips for all Developers: Start by deeply understanding the pain points of your target users. Interview stakeholders, observe workflows, and prioritize addressing the most critical issues to ensure your solution delivers tangible value.
From the outset, Sycai adopted an iterative approach to development. Early MVPs were tested directly in hospital environments, providing immediate feedback from users. For example, radiologist feedback highlighted the need for the tool to integrate seamlessly with existing PACS systems, prompting adjustments to the product's design. This allowed the team to refine their solution, ensuring it aligned with radiologists' needs and integrated smoothly with existing systems.
Key Insight: Direct feedback from end users is invaluable. Engage stakeholders early, test your prototypes in real-world environments, and remain flexible to pivot based on insights.
One of the most significant challenges in healthcare AI development is navigating regulatory requirements. Javier emphasized the importance of considering the regulatory pathway from day one. For instance, Sycai’s team implemented rigorous clinical validation, including comprehensive risk mitigation plans, to achieve CE marking for their product.
Tips to all Developers: Work with regulatory consultants early to align your development processes with compliance requirements. Define and mitigate risks, and prioritize patient outcomes in your design.
High-quality, well-annotated data is the backbone of any AI model. For Sycai, creating a balanced and unbiased dataset involved working with radiologists to annotate lesions and ensuring diversity across parameters such as patient demographics, imaging modalities, and lesion types. Consensus labelling by multiple radiologists was critical to ensuring accuracy.
Lesson Learned: Invest in data curation and annotation. Plan for diversity and balance in your dataset to ensure your model performs well across different scenarios.
Deploying an AI solution in clinical settings introduces additional challenges. Hospitals often have strict requirements for cybersecurity and infrastructure, and many lack cloud connectivity. Sycai optimized its AI models for on-premises deployment, ensuring compatibility with hospital systems such as PACS. For example, they reduced memory usage by optimizing their neural network architecture, aligning with the constraints of on-site servers.
Pro Tip: Anticipate the technical constraints of your end users. Optimize your models for efficiency, and design with scalability and interoperability in mind.
Radiology AI researchers face persistent challenges in clinical deployment, from integration complexity to regulatory compliance. deepcOS® stands out by addressing these challenges through purpose-built solutions that bridge the gap between research and clinical practice. At its core, the platform features a self-service workflow that makes research AI engine submission and deployment transparent and efficient - backed by an industry-leading SDK, comprehensive developer documentation, and seamless MONAI compatibility. By providing sophisticated validation workflows, ready-to-use infrastructure, and flexible deployment options, deepcOS® enables researchers to maintain focus on algorithm development while ensuring robust integration. The platform's built-in governance framework and monitoring tools streamline the technical implementation and compliance processes, supporting the path to publication and commercialization.
Developer Advice: Utilize platforms that provide intuitive integration tools and robust SDKs, compliance support, and monitoring tools to streamline your development process. Learn more at deepci.ai/build.
AI development doesn’t end at deployment. Continuous monitoring and feedback collection are essential to ensure your solution maintains high performance and evolves with user needs. Sycai implemented self-diagnostic features and post-market surveillance plans to identify and address issues proactively. For instance, they tracked error rates and incorporated user feedback into iterative updates.
Key Takeaway: Implement systems for monitoring and feedback. Use analytics to measure the impact of your solution and iterate based on real-world performance.
Bringing an AI solution to life in healthcare is a multidisciplinary effort that requires collaboration, perseverance, and a user-centric approach. As Javier highlighted, flexibility and a willingness to learn from others are crucial to navigating this complex journey.
Are you working on an AI project in healthcare? What challenges have you faced, and what strategies have helped you overcome them? Watch the full webinar to explore these topics in depth.
