The research themes within our AI for medicine research group primarily focus on two main areas: computational medical imaging and artificial intelligence (AI) for Medicine. Our investigations revolve around the following key themes:
1. Fractal analysis for quantifying brain complexity during maturation, aging, and in neurological disorders
Background: The brain exhibits a remarkable degree of statistical self-similarity, and its complexity can be effectively described using fractal geometry. Fractal descriptors provide a valuable means of detecting subtle changes in the brain and monitoring the progression of neurodegenerative conditions. These descriptors capture both strict and statistical self-similarity across a range of spatial scales.
Purpose: Our objective is to design, develop, and apply innovative fractal descriptors that reliably quantify the structural and functional complexity of the brain. We aim to employ these descriptors to study both healthy individuals and patients with neurological disorders. By leveraging fractal analysis, we seek to gain insights into brain maturation, aging processes, and the manifestation of various neurological conditions.
2. AI methods tailored for improving diagnostic accuracy and characterization of neurological disorders
Background: Clinical datasets used in deep learning applications often suffer from limited sample sizes. However, advancements in AI techniques including transfer learning, generative adversarial networks have paved the way to overcome, at least partially, this limitation.
Purpose: Our primary goal is to design, develop, and employ dedicated AI methods that enhance the training and generalization capabilities of deep learning models when working with small-sized datasets in the context of various neurological disorders. By leveraging these specialized AI techniques, we aim to improve the accuracy and precision of diagnostic processes, enabling a more comprehensive characterization of neurological conditions.
3. Overcoming data limitations through generative models for the realization of high-resolution labeled medical imaging datasets
Background: In the realm of deep learning applications, clinical datasets often present challenges such as small sizes and imbalanced data distributions. These limitations can impede the development of accurate and robust models. Generative models, including Generative Adversarial Networks (GANs) and Stable Diffusion, offer the potential for creating realistic synthetic clinical data, addressing data availability limitations.
Purpose: Design and develop high-resolution medical imaging datasets using conditional GANs and diffusion models. Overcome data scarcity with synthetic labeled data, enabling accurate and robust analysis and research in healthcare.
4. Advanced deep learning for risk stratification in lung cancer screening
Background: Lung cancer is a global cause of cancer-related deaths, requiring effective screening methods. The current selection of high-risk individuals based solely on smoking behavior has led to overdiagnosis. Including coronary artery calcifications (CAC) and emphysema scores from LDCT scans, which are associated with cardiovascular disease (CVD) and COPD, the main causes of death alongside lung cancer in LCS cohorts, can improve the risk stratification process, leading to personalization of the screening interventions.
Purpose: Develop and validate a deep learning-based tool that utilizes demographic information, personal history, smoking behavior, and imaging scores (CAC and emphysema) to provide different risk profiles for each level of risk in low-dose CT lung cancer screening and improve the accuracy and effectiveness of lung cancer screening programs.
5. AI for augmented reality in robot-assisted surgery
Background: Three-dimensional models reconstructed from MRI and CT images are useful to guide the surgeon during crucial surgical steps. Recent research activities to improve 3D-guided robotic surgery focus on the development of augmented reality systems able to visualize 3D models in the surgical field with in vivo co-registration during robotic surgery. However, the 3D model does not follow automatically the real organ during its mobilization and traction during surgery. Thus, this non-automated augmented reality may increase the surgical time due to the need of several re-adjustments of the 3D model over the surgical field during the dissection.
Purpose: To design, develop, and validate an AI-based augmented reality system able to automatically co-register, in real time, a digital 3D-model of biological structures and the surgical field.
