Matteo Di Gregorio
Curriculum Learning Methodologies for Deep Learning in Medical Image Analysis.
Rel. Eliana Pastor, Eleonora Poeta. Politecnico di Torino, Corso di laurea magistrale in Ingegneria Informatica (Computer Engineering), 2026
|
Preview |
PDF (Tesi_di_laurea)
- Tesi
Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) | Preview |
Abstract
Deep learning-based segmentation models can exhibit inconsistent performance across patient subgroups and acquisition settings, limiting their reliability when applied to data that differ from the training distribution. Standard training paradigms, such as Empirical Risk Minimization (ERM), primarily optimize overall predictive performance and do not explicitly account for differences in difficulty or performance between subpopulations. As a result, performance disparities may persist for less represented or intrinsically challenging groups. Alternative training strategies, such as Curriculum Learning (CL), have been proposed to improve model optimization by organizing training samples according to their difficulty and progressively exposing the model to more challenging examples. Despite the substantial progress enabled by CL, conventional approaches typically define difficulty using model-derived quantities, such as training loss, which may lack clinical interpretability.
Consequently, the resulting difficulty levels are not necessarily associated with clinically meaningful patient characteristics or acquisition-related factors that systematically influence segmentation performance
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
URI
![]() |
Modifica (riservato agli operatori) |
