Thesis title: Design, development, and implementation of an artificial intelligence-based expert system for diagnosing, monitoring, and evaluating treatment in patients with metastatic melanoma using whole-body FDG-PET/CT images.

Abstract: Metastatic Melanoma (MM) is a very aggressive type of cancer that produces metastases throughout the body with poor survival rates. MM's treatment is very challenging due to its ability to rapidly disseminate to multiple sites and organs and the complex heterogeneous structure of the tumors. In this regard, personalized medicine that supports cancer and treatment monitoring to allow timely adaptation of the therapeutic plan is increasingly important. For cancer monitoring, a combination of two imaging modalities is commonly used, the Positron Emission Tomography (PET) with Computed Tomography (CT). In PET imaging with the tracer 18F-Fluorodeoxyglucose (FDG), tumor regions exhibit high uptake of the tracer. FDG-PET enables clinicians to extract biomarkers crucial for treatment monitoring. However, quantitative measurements require accurate whole-body tumor lesion identification and delineation, while manual segmentation of multiple heterogeneous lesions is labor-intensive, time-consuming and can cause delays in the clinical workflow. Despite the development of semi-automatic and automatic FDG-PET/CT tumor segmentation methods, performance remains limited by false-positive segmentations and missed lesions. Key challenges include variability in tumor morphology, low spatial resolution of PET images and the tracer accumulation in non-tumor regions such as organs and other physiologic uptake. Additionally, the reliance of widely used and highly accurate Deep Learning methods on large annotated datasets is challenging to satisfy in cancer types with limited case numbers, such as MM. This PhD thesis aims to address these issues and support the quantitative evaluation of whole-body 3D FDG-PET/CT images in the management of metastatic melanoma treatment by developing a comprehensive decision support system (DSS).

First, an unsupervised clustering-based method that combined features of the voxels and their neighborhood and an ensemble procedure is developed to segment regions of high uptake from FDG-PET/CT images. Building on this, a comprehensive radiomics analysis led to a tailored pipeline for per-region feature extraction and a deep learning based classification framework that can distinguish true tumor lesions from non-tumor uptake within the detected regions.

Subsequently, a deep learning-based representation learning method is developed to improve feature extraction and learn to extract highly discriminative embeddings of these regions directly from the data. Specifically, a novel position-enhanced learning scheme is proposed to effectively integrate semantic and position-based features, yielding representations useful for downstream clinical tasks, e.g., classification. To this end, a new 3D position-aware Position Encoding Block (PEB) is introduced inside the representation learning method to encode spatial information into the feature vector. The combination of semantic and position information provides enhanced representations where the semantic features are influenced, amplified or suppressed, according to the studied region's anatomical position, size and rotation.

Building on them, the aforementioned algorithms are integrated into a unified pipeline to evaluate a semi-automatic quantitative approach for predicting overall survival. Segmentation masks initially generated from the proposed pipeline and subsequently refined through semi-automatic corrections are used to calculate total metabolic tumor volume (TMTV) and total lesion glycolysis (TLG) from scans acquired at three time points per patient. The longitudinal analysis supported the prognostic value of these Artificial Intelligence (AI)-derived metrics, with significant implications for treatment monitoring and personalized therapy.

Taking into account the scarcity of data in medical imaging datasets, such as those of MM patients, and the recent advances in generative AI, a whole-body FDG-PET synthesis methodology based on diffusion and rectified flow modeling was developed for data augmentation in tumor segmentation. First, a Variational Autoencoder (VAE) model was trained to generate compressed latent representations of 3D PET images and subsequently, a latent lesion-guided PET synthesis model was trained to produce realistic PET images for enhancing the training of deep learning-based segmentation models. Overall, the thesis focuses on the development of AI algorithms to support the clinical workflow in FDG-PET/CT imaging of metastatic melanoma with broader relevance to oncology imaging.

Supervisor: Professor George Matsopoulos

PhD Student: Theodoros P. Vagenas