Βy Nickolas Papanikolaou [Principal Investigator in Oncologic Imaging at the Fundação Champalimaud]
Why did you select prostate cancer as the center of your research? What kind of clinical unmet needs will be addressed in ProCancer-I?
Prostate cancer is among the most prevalent cancers with rising trends, affecting a significant number of patients after the age of 50. Hopefully the vast majority of cases are indolent and only few of them are aggressive that need immediate treatment. With the advent of improved detection methods based on serum PSA, overdiagnosis and overtreatment was the result, leading to the need not only to detect but also to characterize prostate cancer cases according to their biological aggressiveness, and possibly spare invasive and toxic treatment in those patients that they would not benefit due to low aggressive tumors.
How does ProCancer-I differentiate from other EU funded projects? What innovations will it bring to the clinical setting?
ProCancer-I is unique not only due to the comprehensive assessment of unmet clinical needs spanning throughout the whole disease continuum, but also due to the availability of data dealing with the quantity, quality and diversity demands, necessary for proper training and validation of AI models. ProstateNET, the platform to be built collecting available data, is targeting the unparallel number of 1.5 million prostate images, while its existence and maintenance will give unique opportunities to tackle important unsolved problems and create value for our patients, physicians as well as the health care systems.
In ProCancer-I you are using AI approaches to analyze the data. What would be the role of AI – acting autonomous or as a virtual assistant of the end users?
Europe already decided to support and develop human centric AI. That is a decision that we need to serve and adapt our approaches to meet the necessary objectives. The prostate AI models that will be developed from the ProCancer-I consortium will provide valuable data and informed insights to the end-users helping them to decide among different treatment options, weight the benefits and consequences of specific therapeutic schemes. It is our intention and obligation to facilitate the latter by increasing the trustworthiness of the developed models by providing evidence on the model decision mechanisms, abandoning “arrogant” AI approaches presented so far, where AI models have always an opinion no matter how wrong or how right that can be. Concepts like FAIRness, identification of bias, interpretable algorithms are in the DNA of our methodological approaches.
Why is the collection of data and analysis of data a core activity of the ProCAncer-I project?
The size and the diversity of the data will create immense opportunities towards answering important methodological dilemmas that are challenging and relate to whether someone should harmonize and standardize the data working in “laboratory” conditions, or would choose a more “dirty” approach to expose the models in real world data and therefore make them more robust and reproducible. Hybrid approaches using deep and handcrafted features, centralized and federated learning schemes, training from scratch or transfer learning methods in a scale that hasn’t been tried out before will constitute our methodological approaches, expecting that will bring us steps ahead of the current state of the art.