Paletta, Federico Davide
(2026)
Lymphoma segmentation in FDG PET/CT using an Atlas-derived metabolic prior.
[Laurea magistrale], Università di Bologna, Corso di Studio in
Physics [LM-DM270]
Documenti full-text disponibili:
Abstract
FDG PET/CT plays a central role in disease staging and treatment response assessment in lymphoma. Measuring the total volume of lymphoma lesions requires accurate lesion delineation. However, manual lesion segmentation is time-consuming and subject to observer variability. Deep learning methods can automate segmentation, although physiological FDG accumulation makes delineation of tumor challenging. This work aimed to develop and evaluate a pipeline for automatic lymphoma segmentation integrating a population-based metabolic atlas as complementary information. A whole-body atlas was constructed from 50 negative PET/CT scans through CT-based deformable registration, expressed as voxel-wise maps of median FDG uptake and median absolute deviation. These maps were registered to lymphoma patients to derive a prior z-score representing uptake deviations from the population reference. The prior was integrated into a dual encoder nnU-Net architecture, combining PET/CT and prior features at deeper resolution stages. The proposed model and a PET/CT-only baseline were trained using five-fold cross validation on 116 scans and evaluated on 22 patients. The atlas maps reproduced physiologically plausible uptake distributions. The proposed model achieved a mean dice similarity coefficient (DSC) of 0.760, precision of 0.785 and recall of 0.781. The baseline model achieved a mean DSC of 0.778, precision of 0.782 and recall of 0.825. The atlas-derived prior helped reduce false-positive predictions, at the expense of tumor coverage. The developed pipeline demonstrates the feasibility of constructing and integrating a metabolic prior into a segmentation framework. In the evaluated configuration the method did not improve overall performance over PET/CT alone. Further investigation of prior representations and integration strategies, together with evaluation on larger cohorts, is needed to assess the potential of physiological reference information in automatic lymphoma segmentation.
Abstract
FDG PET/CT plays a central role in disease staging and treatment response assessment in lymphoma. Measuring the total volume of lymphoma lesions requires accurate lesion delineation. However, manual lesion segmentation is time-consuming and subject to observer variability. Deep learning methods can automate segmentation, although physiological FDG accumulation makes delineation of tumor challenging. This work aimed to develop and evaluate a pipeline for automatic lymphoma segmentation integrating a population-based metabolic atlas as complementary information. A whole-body atlas was constructed from 50 negative PET/CT scans through CT-based deformable registration, expressed as voxel-wise maps of median FDG uptake and median absolute deviation. These maps were registered to lymphoma patients to derive a prior z-score representing uptake deviations from the population reference. The prior was integrated into a dual encoder nnU-Net architecture, combining PET/CT and prior features at deeper resolution stages. The proposed model and a PET/CT-only baseline were trained using five-fold cross validation on 116 scans and evaluated on 22 patients. The atlas maps reproduced physiologically plausible uptake distributions. The proposed model achieved a mean dice similarity coefficient (DSC) of 0.760, precision of 0.785 and recall of 0.781. The baseline model achieved a mean DSC of 0.778, precision of 0.782 and recall of 0.825. The atlas-derived prior helped reduce false-positive predictions, at the expense of tumor coverage. The developed pipeline demonstrates the feasibility of constructing and integrating a metabolic prior into a segmentation framework. In the evaluated configuration the method did not improve overall performance over PET/CT alone. Further investigation of prior representations and integration strategies, together with evaluation on larger cohorts, is needed to assess the potential of physiological reference information in automatic lymphoma segmentation.
Tipologia del documento
Tesi di laurea
(Laurea magistrale)
Autore della tesi
Paletta, Federico Davide
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Indirizzo
Applied Physics
Ordinamento Cds
DM270
Parole chiave
Lymphoma,Segmentation,Deep learning,nnUNet,TMTV,Atlas,Deformable registration
Data di discussione della Tesi
29 Ottobre 2026
URI
Altri metadati
Tipologia del documento
Tesi di laurea
(NON SPECIFICATO)
Autore della tesi
Paletta, Federico Davide
Relatore della tesi
Correlatore della tesi
Scuola
Corso di studio
Indirizzo
Applied Physics
Ordinamento Cds
DM270
Parole chiave
Lymphoma,Segmentation,Deep learning,nnUNet,TMTV,Atlas,Deformable registration
Data di discussione della Tesi
29 Ottobre 2026
URI
Statistica sui download
Gestione del documento: