Διδακτορικές Διατριβές

Μόνιμο URI για αυτήν τη συλλογήhttps://repository.ihu.gr/handle/11544/47648

Περιηγούμαι

Πλοήγηση Διδακτορικές Διατριβές ανά Θέμα "Melanoma thickness prediction"

A B C D E F G H I J K L M N O P Q R S T U V W X Y Z
Α Β Γ Δ Ε Ζ Η Θ Ι Κ Λ Μ Ν Ξ Ο Π Ρ Σ Τ Υ Φ Χ Ψ Ω
Φιλτράρετε τα αποτελέσματα πληκτρολογώντας τα πρώτα γράμματα
Τώρα δείχνει 1 - 1 από 1
  • Μικρογραφία εικόνας
    Τεκμήριο
    Holistic deep learning approaches on medical signals/images
    (ΔΙΠΑΕ, 2026-07-28) Ntampakis, Nikolaos; Diamantaras, Konstantinos; Σχολή Μηχανικών, Τμήμα Μηχανικών Πληροφορικής και Ηλεκτρονικών Συστημάτων
    This dissertation develops and validates a holistic deep learning framework for the analysis of medical signals and images, structured around three sequentially dependent stages: data quality assurance, deep learning model development, and clinical interpretability. The structure responds to a methodological gap in contemporary medical AI research, in which contributions typically address a single clinical problem, operate on a single data modality, and optimize a single performance metric within a single pipeline stage. The proposed structure aligns with the operational dimensions articulated by the FUTURE-AI international guideline, the requirements of the EU AI Act, and the World Health Organization guidance on AI for health. The contribution is vertical rather than horizontal: the dissertation constructs each pillar to a depth of methodological rigor that existing consensus frameworks identify as necessary but do not themselves construct, and demonstrates the resulting pipeline through six peer-reviewed contributions that converge on clinical domains, mainly dementia and melanoma, each traversed end-to-end from data preparation through modelling to explanation. Stage A introduces SPINE, a multi-criteria post-generation evaluation methodology for synthetic medical images integrating expert domain assessment, statistical data analysis, and adversarial evaluation. Applied to chest X-ray images synthesized through a conditional GAN, the sequential application of quality criteria produced a cumulative improvement in structural similarity between the synthetic and real distributions, with the framework designed to be modality-agnostic and extensible to other modalities. Stage B comprises four deep learning applications distributed across the clinical domains of dementia and melanoma. For dementia, three complementary contributions approach the condition through distinct sensory channels: an ML classifier trained on longitudinal records predicts future onset in initially healthy individuals with a perfect macro-F1 on the test set, with explainable AI feature attributions consistent with established clinical risk factors; the DEFORMISE framework combines anatomically grounded edge-based MRI slice selection with a confidence-based committee of three architecturally heterogeneous 3D networks, achieving state-of-the-art accuracy on the combined OASIS-1 and OASIS-2 test set, with cross-dataset generalization confirmed on external ADNI patients; and the NeuroXVocal system fuses acoustic features, speech embeddings, and textual representations within a transformer-based three-stream classifier, surpassing all prior published methods on the ADReSSo benchmark, while its RAG-based explainer attained a low misinterpretation risk score under structured clinical evaluation by medical professionals. For melanoma, a DNN classifier coupled with a lesion isolation preprocessing pipeline reached state-of-the-art accuracy and macro-F1 on three-class thickness prediction over the Derm7pt benchmark. Stage C presents Fusion Grad-CAM, a methodology for generating consolidated attention maps for majority voting ensemble classifiers through confidence-weighted fusion restricted to majority-contributing models, validated across two-, three-, and four-model ensembles and directly applicable to the Stage B committee architectures. The methodological connective tissue, Stage A on the data side and Stage C on the explanation side, is designed to extend onto the Stage B modalities, with the SPINE framework directly transferable to dermoscopic and MRI inputs and Fusion Grad-CAM directly applicable to the DEFORMISE committee. Taken together, the contributions support the central claim of the dissertation: that the lifecycle structure prescribed by horizontal governance frameworks constitutes a technically achievable engineering paradigm when each pillar is vertically instantiated with domain-specific methodological rigor.