S2450
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Conclusion: The trained model demonstrates the feasibility of using deep learning algorithms to perform reliable CT- to-PET mapping for the head and neck anatomical region, capable of generalized CT-to-PET conversion and not just merely acting as a proxy for FDG uptake detection on a CT acquired within a PET scanner. Both model architecture and weights are publicly available on GitHub [5]. References: 1. M. Salehjahromi et al. Synthetic PET from CT improves diagnosis and prognosis for lung cancer: Proof of concept. Cell reports. Medicine, 5(3), 101463. https://doi.org/10.1016/j.xcrm.2024.1014632. P. Isola et al. Image-to-image translation with conditional adversarial networks.arXiv preprint arXiv:1611.07004.https://arxiv.org/abs/1611.0 70043. K. He et al. Deep residual learning for image recognition. arXiv.https://arxiv.org/abs/1512.033854. https://heckto r.grand- challenge.org5. https://github.com/maksymfritsak/Hea d-and-Neck-sPET Keywords: Synthetic PET, Head and Neck Cancer, FDG-PET/CT Digital Poster Highlight 2770 Microstructural Parameters from Diffusion- Weighted MRI via Machine Learning and realistic in silico cellular models for Breast and Prostate Cancer Chiara Tinelli, Chiara Scotti, Fabio Casaccio, Guido
carcinoma, some of which comprise additional subtypes. DW signals were simulated using a GPU- accelerated MC method that models water diffusion and microenvironment interactions [4]. By leveraging the constructed library, a ML framework capable of extracting voxel-wise quantitative descriptors of the tissue microenvironment (i.e. mean cellular volume, diffusion coefficient and apparent cellularity ) from conventional DWI data was implemented. The proposed architecture is a multi-layer perceptron (MLP) with an input layer including three b-value nodes and three corresponding normalized DW signal nodes, followed by fully connected hidden layers. The proposed approach has been applied to in vivo DWI of 50 breast cancer patients [6] and 50 prostate cancer patients [7] obtained from The Cancer Imaging Archive. The extracted microstructural maps were exploited for patient histological stratification (Mann- Whitney test, α =0.05) according to Scarff-Bloom- Richardson (SBR) and Gleason Score for breast and prostate cancer patients, respectively. Results: Histogram-based analysis of the extracted microstructural maps on the GTV of breast cancer tumors (Figure 1) revealed statistically significant differences between low and intermediate/high SBR grades (Figure 2). Additionally, the parameters extracted from DWI data of prostate cancer patients revealed a similar pattern in stratifying tumors according to Gleason Score. The stratification performance was superior compared to both a ML framework trained on literature-based non-realistic substrates [2] and conventional Apparent Diffusion Coefficient-based characterization (Figure 2.a).
Baroni, Letizia Morelli, Chiara Paganelli DEIB, Politecnico di Milano, Milano, Italy
Purpose/Objective: Given the invasiveness and limited sampling of biopsies, Diffusion-Weighted MRI (DWI) is considered a non-invasive modality to assess tumor microstructure heterogeneity [1]. Recent DWI-based analytical models estimate sub-voxel microstructural properties by mapping in vivo DW signals to simulated dictionaries, but are limited by non-realistic tissue models and the need for re-optimization with changing acquisition parameters [2], [3]. This work aims at developing a generalizable Machine Learning (ML) algorithm to extract microstructural parameters starting from a library of in silico realistic cellular substrates and corresponding Monte Carlo simulated DW signals [4]. Material/Methods: Microscopy images from the LIVECell dataset [5] were used to generate in silico substrates dynamically grown via a Monte Carlo (MC) model. The cell lines considered include three main types, i.e. breast adenocarcinoma, ovarian adenocarcinoma and liver
Made with FlippingBook - Share PDF online