In a collaborative effort with the Champalimaud Foundation (CF), the Faculty of Sciences at the University of Lisbon and the University of Coimbra, the NOVA Analytics Lab from NOVA IMS, led by Leonardo Vanneschi, developed an AI-based system with predictive capabilities in oncology. The goal was to leverage artificial intelligence (AI) techniques to enhance radiomics by extracting informative features from available medical images, using precise and robust predictive models.
The newly developed AI techniques from NOVA IMS successfully unveiled previously unknown features that distinguish breast and axillary cancers, significantly improving early diagnostic accuracy and contributing to a more tailored diagnostic approach in clinical settings. Notably, the model achieved an unprecedented 87% accuracy in predicting the complete pathological response of axillary tumors.
This work also had a major impact on the segmentation and detection of prostate cancer, utilizing the diverse ProstateNet dataset from CF, which includes over 1,000 samples from three different scanners and 14 institutions, along with the integration of three other independent datasets. The study demonstrated that models trained on large, diverse datasets generalize better, making them more reliable for lesion detection.
Breast and prostate cancers are among the most prevalent types of cancer worldwide. Additionally, axillary cancer, often stemming from complications related to breast cancer, has a significant mortality rate. In most cases, this mortality is due to the late detection of the disease, as many patients remain asymptomatic for long periods. This situation underscores the urgent need to develop reliable predictive models that can enhance personalized therapies by estimating individual treatment responses and forecasting risks associated with tumor lesions. The solution developed by the NOVA team aims to tackle these challenges and overcome limitations of existing models that are often based on limited datasets (usually from a single institution and patient cohort) and rely on pre-existing AI algorithms, restricting their general applicability.
The algorithms and models developed by the NOVA Analytics Lab were integrated into an innovative software system that has been validated by clinicians at CF and incorporated into their existing framework. The system is in function since approximately three years, and it was used so far for research purposes, with the perspective of using it on patients in the future. This cutting-edge AI-driven system has already shown significant advancements in cancer diagnosis and treatment, leading to improved early detection rates and more accurate predictions of patient responses to therapies.
This major project was a collaborative effort, combining powerful computational infrastructure with exceptional talent, resulting in two PhD dissertations, eight master’s theses, 17 articles, one book chapter, and numerous presentations at international conferences. As a result of this research at NOVA IMS, oncology clinicians at the Champalimaud Foundation have the potential to design more personalized treatment plans while also reducing costs for the healthcare system.
Ultimately, the true beneficiaries of this research will be the patients, who will be able to receive earlier, more accurate, and less invasive diagnoses, greatly improving their treatment outcomes with fewer side effects and better chances of tumor clearance and survival. This work has the potential to be implemented in medical institutions globally and extended to other tumor types, unlocking access to its benefits for millions of people worldwide.
The objective for the future is to establish a vast network of medical institutions to share data, as larger and more diverse data will lead to improved accuracy in the algorithms.
L. Vanneschi