Advancing AI in Healthcare: ITI’s Journey to Revolutionize Breast Cancer Detection
The Healthcare Data Innovation Council selected this story from ITI for its efforts to develop advanced tools for early detection and prognosis of breast cancer, showcasing how to enhance medical outcomes. This talks about the technical achievements, practical challenges, and regulatory hurdles faced in bringing such innovations to market, offering valuable insights into the future of AI in healthcare.
The Instituto Tecnologico de Informatica (ITI), based in Valencia, Spain has been working for several years on Big Data and Artificial Intelligence to improve the prevention, diagnosis, and treatment of chronic diseases by developing advanced tools and predictive models.
One of their key projects focuses on automating the analysis of medical images to aid in breast cancer screening, which is the most common among women in Spain. Although the mortality rate is decreasing thanks to medical advances, the incidence continues to rise. Early detection systems for breast cancer require many hours of radiologists evaluating digital mammograms for signs of the disease. Automatic detection would substantially reduce specialists’ dedication to this task, contributing to faster attention in radiology services. This work has significant potential in international markets, particularly for dense breast tissue common among certain populations. It is currently at the clinical validation stage in Spain and it aims to produce a certified medical device that could embed the created automation.
We explored some of the hurdles encountered by ITI and typical in this field, from the technical to the commercial point of view.
The first common issue is related to data provision, since in Spain, like other countries in Europe, the mammograms belong to the public health and not to the hospitals. Other datasets could be used but often only for research purposes or, if for commercial purposes, under a payment in percentage of the outcomes and earnings. For ITI, these data availability issues ended up doubling the time for the development of their tool, thus slowing the adoption of a potentially life-saving solution.
A second issue related to the datasets was that they contained biases. In the case of mammograms, these biases arise from the use of different X-ray machines across various institutes, leading to variations in data formats and resolutions. Harmonizing this data through machine learning techniques was essential.
Ensuring high data quality is crucial when developing AI solutions.
When addressing issues related to commercialization, a common problem is that not all diagnostic and prognostic tools with life-saving potential are commercially viable. To sustain their development, companies like ITI must focus on diseases with a higher potential for financial return. Alternatively, they need to find commercial partners who can sponsor the development of these tools. This often means prioritizing conditions that are more prevalent or have a higher demand for advanced diagnostics, ensuring that the resources invested can be recouped
Thirdly, bringing a medical device to market depends significantly on its classification and the associated regulatory requirements and it could take from 2-3 extra years of work for class I devices to four times more for IIb or III classes. Moreover, it requires experts in cybersecurity and usability. In fact, the AI Act delegates the responsibility for regulating AI within the medical device industry. This means that medical devices incorporating AI will need to be audited by experts from notified bodies who can assess the risk management of these systems. Additionally, unlike traditional medical devices, AI systems require continuous monitoring and management due to their evolving nature. This includes ongoing risk management, tracking data biases, and addressing model drift. Ethical considerations and AI ethics management must be maintained throughout the lifecycle of the device, from development to decommissioning. So, transitioning from a research model to a regulatory-compliant product takes a long time and increases the cost of development.
ITI did excellent work dealing with all these challenges and taking the big step of transitioning from research to market.
A take-home message we learn from them is to keep in mind the difference between research and product development: when doing research, there is no need to get tangled up in regulations. The advice is to speed up the research as fast as possible with the data available for research. Even though the data may not be used in the market, it will provide a screenshot of what is achievable. Only then can the product development journey start, and there will be different challenges because the model will not perform as well on different data, so it might have to be adapted.
It is exciting to see initiatives like ITI’s progressing in the stages of AI medical devices. In the course of the next decade, we expect to see many hospital departments employing AI diagnostics tools to support medical decisions.