Beyond supervised deep learning frameworks for precision medicine

Espis, Andrea (2026) Beyond supervised deep learning frameworks for precision medicine, [Dissertation thesis], Alma Mater Studiorum Università di Bologna. Dottorato di ricerca in Scienze e tecnologie della salute, 38 Ciclo.
Documenti full-text disponibili:
[thumbnail of Revised_Andrea_Espis_Thesis_20260624.pdf] Documento PDF (English) - Richiede un lettore di PDF come Xpdf o Adobe Acrobat Reader
Disponibile con Licenza: Salvo eventuali più ampie autorizzazioni dell'autore, la tesi può essere liberamente consultata e può essere effettuato il salvataggio e la stampa di una copia per fini strettamente personali di studio, di ricerca e di insegnamento, con espresso divieto di qualunque utilizzo direttamente o indirettamente commerciale. Ogni altro diritto sul materiale è riservato.
Download (11MB)

Abstract

Deep learning has achieved strong results in medical image analysis, yet clinical translation is constrained by the realities of routine practice: limited and weak labels, small and imbalanced cohorts, heterogeneous acquisition pipelines, and distribution shift. This thesis investigates how to strengthen the methodological foundations of medical AI under these constraints, moving beyond standard supervised learning and adopting evaluation protocols for reproducible, clinically oriented conclusions. The same focus spans two settings: in radiology and ophthalmology (brain MRI, chest X-ray, retinal OCT), images are analyzed as single inputs; in computational pathology, gigapixel whole-slide images (WSIs) are analyzed through patch-based models, where weak supervision, reproducibility at scale, and robustness to acquisition shift become central. In the first setting, controlled experiments show that comparisons between supervised (SL) and self-supervised learning (SSL) are often confounded by non-equivalent workflows; under matched workflows on small, imbalanced datasets, domain-agnostic SSL does not consistently outperform strong supervised baselines, with clear gains only at much larger pre-training scale. To improve data efficiency, MedBooster builds domain-specific SSL pretext tasks from programmatically extracted features; its instantiation, NeuroBooster, matches or surpasses these baselines on brain-MRI age prediction and Alzheimer's-disease classification, most clearly in low-label regimes. In computational pathology, a tumor-site classification pipeline is reproduced on open-access TCGA data, and learning curves reveal tissue-specific data requirements; under matched conditions, a pathology foundation encoder improves data efficiency over a natural-image-pretrained one. For resolution-driven shift (microns-per-pixel), metadata-driven patch harmonization improves out-of-distribution performance while preserving in-distribution accuracy. A multimodal classifier combining WSI representations with mutation-status features shows that each modality's usefulness depends on the evaluation setting, motivating fusion strategies sensitive to modality reliability. Overall, progress depends not only on adopting paradigms beyond supervised learning, but on the conditions of their application and evaluation: pre-training scale and relevance, reproducible evaluation, and designs targeting distribution shift in clinical translation.

Abstract
Tipologia del documento
Tesi di dottorato
Autore
Espis, Andrea
Supervisore
Co-supervisore
Dottorato di ricerca
Ciclo
38
Coordinatore
Settore disciplinare
Settore concorsuale
Parole chiave
Medical image analysis, computational pathology, whole-slide images, cancer, neuroimaging, self-supervised learning, multiple instance learning, multimodal learning, generalization, reproducibility
Data di discussione
15 Luglio 2026
URI

Altri metadati

Statistica sui download

Gestione del documento: Visualizza la tesi

^