Alex Ferroni
Self-Supervised Pretraining and Transfer Learning for Biomedical Classification from Raman Data.
Rel. Lia Morra. 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
Raman spectroscopy provides label-free molecular fingerprints of biological samples and is therefore a promising tool for rapid biomedical diagnosis. However, deep learning applications in this field are limited by the scarcity of large annotated datasets, since reliable clinical labels are expensive and difficult to obtain. This thesis investigates self-supervised pre-training as a strategy to learn transferable Raman representations from unlabeled spectra and reduce the dependence on task specific labels. Two complementary self-supervised approaches were studied. Barlow Twins was adapted to one-dimensional Raman spectra to learn representations invariant to realistic perturbations such as noise, scaling, and baseline variations. In parallel, a Spectral Masked Autoencoder (SMAE) was implemented to reconstruct masked spectral regions and capture local-to-global biochemical structure.
The resulting pre-trained encoders were transferred to downstream biomedical classification tasks, including inflammatory bowel disease plasma spectra and BACTERIA-ID pathogen identification
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Ente in cotutela
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
