The study analyzes the role of deep learning in predicting iodine maps from CT images acquired using SECT technology, with the goal of enabling the extraction of valuable clinical information from images even when the most advanced DECT technologies are not available in the facility. A ResNeXt-32 neural network model was developed and trained on a dataset of 115,000 images from 150 different patients, with the dataset divided into 70-20-10% for training, validation, and testing, respectively. The model’s performance was evaluated using quantitative metrics such as Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR), achieving an average PSNR of 23.82 dB. Comparison with the literature confirms the clinical potential of the proposed model, while also highlighting the need for further optimization to improve the resolution of small-caliber vascular structures and the accuracy of iodine quantification.
Lo studio analizza il ruolo del deep learning nella predizione delle mappe di iodio a partire da immagini TC acquisite con tecnica SECT, al fine di consentire l’estrazione di importanti informazioni cliniche provenienti dalle immagini anche quando non sono presenti le più avanzate tecnologie DECT nella struttura. È stato sviluppato e addestrato un modello di rete neurale ResNeXt-32 su un campione di 115.000 immagini provenienti da 150 pazienti diversi, con suddivisione del data-set in 70-20-10% (training, validation e test). Le prestazioni del modello sono state valutate tramite metriche quantitative quali Mean Squared Error (MSE) e Peak Signal to Noise Ratio (PSNR) ottenendo un PSNR medio di 23.82 dB. Dal confronto con la letteratura si conferma la potenzialità clinica del modello proposto ma si evidenziano anche ulteriori ottimizzazioni per migliorare la risoluzione delle strutture vascolari di piccolo calibro e la quantificazione dello iodio.
Ruolo del Deep Learning nella predizione delle mappe di iodio dalle immagini TC a singola energia
CARBONE, ANDREA
2024/2025
Abstract
The study analyzes the role of deep learning in predicting iodine maps from CT images acquired using SECT technology, with the goal of enabling the extraction of valuable clinical information from images even when the most advanced DECT technologies are not available in the facility. A ResNeXt-32 neural network model was developed and trained on a dataset of 115,000 images from 150 different patients, with the dataset divided into 70-20-10% for training, validation, and testing, respectively. The model’s performance was evaluated using quantitative metrics such as Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR), achieving an average PSNR of 23.82 dB. Comparison with the literature confirms the clinical potential of the proposed model, while also highlighting the need for further optimization to improve the resolution of small-caliber vascular structures and the accuracy of iodine quantification.I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.12075/23888