ESTRO 2025 - Abstract Book

S3022

Physics - Image acquisition and processing

ESTRO 2025

2642

Digital Poster Applying transfer learning from high-energy physics to enhance CT image quality in lung cancer screening Lorenzo Cederle 1 , Francesca Camagni 1 , Mariagrazia Monteleone 1 , Federico Camponovo 2 , Guido Baroni 1 , Pietro Govoni 2,3 , Simone Gennai 2 , Chiara Paganelli 1 1 Department of Electronics, Information and Bioengineering (DEIB), Politecnico di Milano, Milan, Italy. 2 Sezione di Milano - Bicocca, INFN, Milan, Italy. 3 Department of Physics "Giuseppe Occhialini", Università degli Studi di Milano - Bicocca, Milan, Italy Purpose/Objective: This study investigates the feasibility of leveraging high-energy physics (HEP) simulations to train deep learning models for medical imaging tasks. Images from Quantum Chromodynamics (QCD) jets, generated via Monte Carlo simulations, provide a valuable dataset due to their large volume and precise causality relationships among underlying variables. Here, we aim to demonstrate how a generative model pre-trained on HEP images can transfer knowledge effectively for CT image quality enhancement in lung cancer screening. Material/Methods: A pix2pix generative adversarial network (GAN) [1] was initially trained for 100 epochs on 3283 QCD jets images [2], divided into train, validation, and test datasets with 80% train/test ratio (jets model). Such images were produced through Monte Carlo simulations using the Geant4 simulation toolkit [3] and pre-processed as to better resemble CT images, while maintaining their physical interpretability. Different models were then trained on a set of lung CT images from the National Lung Screening Trial (NLST) [4] dataset, reconstructed with two different filters, which constitute the input and target domains, respectively. Specifically, (i) a model was trained from scratch on the NLST data with the same sample size and the same number of epochs as the QCD dataset (full model); (ii) a model was trained from scratch for 20 epochs on 300 NLST images (limited model); (iii) fine-tuning was applied to the HEP-trained network on the small subset of 300 NLST images for 20 epochs, adapting the pre-trained model to the new domain (fine-tuned model). Performance of all networks was evaluated through the following image quality metrics: structural similarity index measure (SSIM), peak signal-to noise ratio (PSNR), root mean square error (RMSE), edge preservation ratio (EPR), and edge generation ratio (EGR) [5]. Results: The fine-tuned model showed better performance than the limited model, with image quality comparable to the full model and to other results found in the literature [6]. Figure 1 shows examples of predictions made by the jets and fine-tuned models, compared to input and target images. Average metrics evaluated on the test datasets for the networks trained on NLST data are reported in figure 2.