S2472
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Keywords: NSCLC, Survival Analysis, Machine Learning
Results:
Digital Poster 3823 Intratumor variation in imaging and histological phenotypes of prostate cancer Erik Nilsson 1 , Josefine Grefve 1 , Joakim Jonsson 1 , Kristina Sandgren 1 , Angsana K Lindberg 1 , Karin Söderkvist 2 , Camilla T Karlsson 2 , Sara Strandberg 3 , Katrine Riklund 3 , Anders Bergh 4 , Andreas Josefsson 5 , Tufve Nyholm 1 1 Department of Diagnostics and Intervention, Radiation Physics, Umeå University, Umeå, Sweden. 2 Department of Diagnostics and Intervention, Oncology, Umeå University, Umeå, Sweden. 3 Department of Diagnostics and Intervention, Diagnostic Radiology, Umeå University, Umeå, Sweden. 4 Department of Medical Biosciences, Pathology, Umeå University, Umeå, Sweden. 5 Department of Diagnostics and Intervention, Urology and Andrology, Umeå University, Umeå, Sweden Purpose/Objective: To characterize the intratumor histological heterogeneity of prostate cancer (PCa) lesions, by comparing immunohistochemical (IHC) expression across regions stratified by qualitative image assessment Material/Methods: Thirty patients with high- or intermediate-risk PCa underwent [68Ga]prostate-specific membrane antigen (PSMA) positron emission tomography (PET)/multiparametric magnetic resonance imaging (mpMRI) before radical prostatectomy. Whole-mount histopathological evaluation provided Gleason grading and immunoreactivity H-scores based on IHC expression of PSMA, tissue prostate-specific antigen (PSA) and the marker of proliferation (Ki-67), providing independent information on tumor aggressiveness1, 2. Immunoreactivity for PSMA and Ki-67 were quantified by the percentage of positive cells. The H- score for PSA was formed by multiplying the staining intensity (0: negative, 1: weak, 2: moderate and 3: intense) by the distribution of positive cells (0: no staining, 1: 1%–25%, 2: 26%–50%, 3: 51%–75%, and 4: 76%–100%). Histopathological information was co- registered with in vivo image data, and regions within histologically confirmed disease were considered positive or negative on PSMA-PET and mpMRI. Regions were further classified as inside or outside the gross tumor volume (GTV), delineated by an experienced radiation oncologist on PSMA-PET/mpMRI (Figure 1). The Mann-Whitney U test and one-sided binomial tests were used for unpaired and paired comparisons, respectively. All p-values < 0.05 were considered statistically significant.
The clinical-only model (CLIN) achieved the strongest standalone performance (C-index = 0.535 ± 0.028, p = 0.025) and remained robust on the external test set (C- index = 0.572, p = 0.099). Among dosimetric sets, DOSE (0.541 ± 0.041, p = 0.045) and EQD2 (0.539 ± 0.035, p = 0.034) outperformed DVH (0.530 ± 0.044) and CT-based features (0.506 ± 0.018).The best- performing multimodal models combined clinical and dosiomics information. The CLIN + EQD2 model achieved a C-index of 0.547 ± 0.041 (p = 0.031) with significant external validation (C-index = 0.559, log- rank p = 0.0225). Adding CT radiomics reduced performance across all combinations, suggesting limited incremental value for OS prediction.Permutation-based feature importance analysis identified age, smoking status, RT technique (3D-CRT vs IMRT), gender, histology, and concurrent chemotherapy as key clinical predictors, while GTV sphericity, GTV entropy, and PTV elongation were dominant EQD2-based dosiomics features. Conclusion: Clinical factors remain the most powerful predictors of overall survival in NSCLC radiotherapy, while EQD2- based dosiomics added complementary prognostic information. CT-based radiomics did not improve performance. The heterogeneity between the RTOG (stage III only) and REQUITE (stages I–III) cohorts, as well as the missing staging, likely limited the model's generalizability and performance. Overall prognostic assessment was challenging. References: [1] Bradley, J.; Forster, K. (2018). Data from NSCLC- Cetuximab. The Cancer Imaging Archive. DOI: http://doi.org/10.7937/TCIA.2018.jze75u7v [2] Seibold, Petra, et al. "REQUITE: a prospective multicentre cohort study of patients undergoing radiotherapy for breast, lung or prostate cancer." Radiotherapy and Oncology 138 (2019): 59-67.[3] Zwanenburg, Alex, et al. "The image biomarker standardization initiative: standardized quantitative radiomics for high- throughput image-based phenotyping." Radiology 295.2 (2020): 328-338.[4] Leo Breiman. Random forests. Machine learning, 45(1):5–32, 2001.
Made with FlippingBook - Share PDF online