RPS 2216 - Structured reporting, radiomics and deep learning

Author Block: M. Mancino, G. Avesani, A. Infante, S. Gaudino, B. Merlino, L. Natale, E. Sala; Rome/IT
Purpose: To evaluate the impact of structured radiology reports (SRRs) on clinical decision-making and patient management, specifically focusing on improvements in diagnostic accuracy, treatment planning, patient outcomes, and data standardization. Additionally, we aim to measure clinicians' satisfaction with the clarity, comprehensiveness, and utility of SRRs in comparison to standard narrative reports.
Methods or Background: SRRs were introduced one and a half years ago at Fondazione Policlinico Universitario A.Gemelli, a large tertiary university italian hospital , and have since been implemented in nearly all radiological procedures . This transition was the result of a comprehensive collaboration among radiologists, surgeons, and clinicians to ensure that the structured templates met clinical needs while adhering to guidelines from major radiological societies. An extensive and anonymous survey was conducted among non-radiologist clinicians, surgeons, and residents from various specialties to gather feedback on SRRs' clarity, clinical impact, adaptability, research value, and efficiency compared to traditional reports.
Results or Findings: Survey responses indicate increased clinician satisfaction, improved communication, and greater workflow efficiency with SRRs. Preliminary findings suggest better data interpretation, ease of retrieval, and enhanced multidisciplinary discussions, particularly in oncology. SRRs were perceived as more effective than traditional reports in supporting clinical decisions and improving collaboration.
Conclusion: SRRs significantly improve clinical practice by delivering clearer, more consistent interpretations that directly enhance patient outcomes and support effective decision-making. The increased clarity and standardization foster better collaboration among clinicians, ultimately benefiting patient care. Continuous feedback from users is essential to refine SRRs and ensure they remain adaptable and impactful. Furthermore, SRRs provide a foundation for consistent data collection, crucial for advancing research and supporting evidence-based practices.
Limitations: The single-center survey limits the generalizability of the results. Future studies should include multiple centers to validate these findings.
Funding for this study: No specific funding was obtained for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study has been notified to the Ethics Committee of Fondazione Policlinico Universitario A. Gemelli IRCCS.

Author Block: D. I. Rodríguez Sánchez , R. Spaans, S. Rostami, O. Maxouri, Z. Bodalal, R. G. H. Beets-Tan; Amsterdam/NL
Purpose: Tumour genetic heterogeneity is a fact in cancer research. While biopsying every lesion is clinically infeasible, imaging-based methods (such as radiogenomics) promise clinicians non-invasive insights into the underlying tumour biology. This study assesses the impact of accounting for tumour heterogeneity by comparing Single Instance Learning (SIL) and Multiple Instance Learning (MIL) AI approaches.
Methods or Background: A cohort of 1666 routine contrast-enhanced CT scans, including over 11.000 segmented lesions, was retrospectively collected along with matched next-generation sequencing data. The morphological phenotype was quantified from each lesion by radiomic features, with subsequent feature selection using orthogonal principal feature selection (OPFA) and five-fold cross-validation. SIL and MIL-based machine learning methods were compared to predict the mutational status of TP53, KRAS, and EGFR. Interlesional morphological heterogeneity was measured using spatial distance metrics.
Results or Findings: For genes with established low biological variability between lesions (TP53 and KRAS), accounting for tumour heterogeneity did not improve the radiogenomic predictive performance. However, for genes with high discordance (EGFR), MIL-based machine learning methods (AUC range=0.72-0.78) performed significantly better than SIL (AUC=0.56). These findings are also supported by interlesional heterogeneity scores, which did not differ between wild-type and mutated TP53 or KRAS cases. Conversely, EGFR-mutated patients demonstrated significantly higher interlesional morphological heterogeneity than their wild-type counterparts (p<0.0001).
Conclusion: MIL models may better reflect tumour heterogeneity, particularly in cases with high biological variability. Incorporating MIL into radiogenomic models may enhance their predictive accuracy by accounting for real-world tumour heterogeneity.
Limitations: The study's limitations include external validation. A large-scale multicenter radiogenomics dataset is currently being finalised to validate these results.
Funding for this study: Funding was provided by the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement number 101034290 (EMERALD International PhD Program for Medical Doctors).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB approval: IRBd19-147

Author Block: S. Kuckertz¹, S. Heldmann¹, F. Peisen², J. H. Moltz³; ¹Lübeck/DE, ²Tübingen/DE, ³Bremen/DE
Purpose: Continuous lesion assessment in cancer patients is integral to radiologists’ work. Part of this process is the tedious and time-consuming (re-)localisation and measurement of lesions. Fast and precise image registration facilitates the workflow by automatically matching previous and current observations.
In this study we evaluate our deep-learning-based lung registration using a longitudinal dataset including expert lesion segmentations, comparing it to a state-of-the-art conventional non-learning approach.
Methods or Background: We follow a 3-level coarse-to-fine deep-learning registration approach. At each level, we input the baseline and follow-up CT scan at a different resolution into a U-Net, resulting in a deformation field that maps corresponding locations from baseline to follow-up. Combining the deformations from all levels allows accurate tracking of anatomical points.
Our method is trained on 681 follow-up image pairs and evaluated on a distinct dataset consisting of 90 image pairs with 307 manually segmented lung lesions. We compare our approach to a non-learning GPU-accelerated registration. Each lesion centre in the baseline scan is propagated onto the follow-up, where we check whether it maps inside the given corresponding lesion.
Results or Findings: With our learning-based approach, 81.1% of the baseline lesion centres were correctly mapped to the corresponding lesion in the follow-up (conventional approach: 73.6%). The median distance between the propagated and the given lesion centre was 1.9 mm (conventional approach: 3.2 mm) and the average calculation time was 0.92 s (conventional approach: 14.56 s).
Conclusion: Our learning-based registration approach enhances both speed and accuracy, enabling precise relocation of all lung lesions in follow-up scans in less than a second. This facilitates radiologists’ workflow, also enabling cursor synchronisation and change highlighting in viewers.
Limitations: CTs were cropped to the thorax area and resampled to an isotropic resolution of 1.5 mm.
Funding for this study: Funding was provided by the Federal Ministry of Education and Research of Germany (BMBF) as part of SPIRABENE (project number 13GW0561B).
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: Only retrospective data was used for this work. The outcome had no effect on patient treatment.

Author Block: A. Ammirabile¹, F. Fiz², E. M. Ragaini¹, S. Sirchia³, S. Viganò¹, M. Francone¹, L. Cavinato¹, E. Lanzarone³, L. Viganò¹; ¹Milan/IT, ²Genova/IT, ³Bergamo/IT
Purpose: To investigate the variation of three textural features (mean HU, entropy, and uniformity) in the peritumoral tissue around colorectal liver metastases (CRLM) as distance from the tumor increases.
Methods or Background: This retrospective study included all consecutive patients with histologically proven CRLM who underwent locoregional treatment (resection/ablation) between January 2010 and December 2022. Inclusion criteria were high-quality CT with an adequate portal phase and identifiable hypodense CRLM (>10 mm). Multiple VOIs were generated: 1) manual tumor segmentation (Tumor-VOI); 2) multiple automatic concentric rims at increasing distance from CRLM (1 to 10 millimeters); 3) manual segmentation of a virtual biopsy of non-tumoral parenchyma (Liver-VOI). Radiomic features were extracted by the LifeX software. The percentage variation of index values across different VOIs was calculated, using Liver-VOI as reference. Subgroup analyses were based on tumor size (10-30 vs. >30 mm) and chemotherapy administration (no chemotherapy vs. responders).
Results or Findings: 63 CRLM in 51 patients (median age 67 years, 14 females) were analyzed. Median peritumoral HU values were similar to Liver-VOI, except within the 1-mm VOI around CRLM (p=0.002). Entropy progressively decreased (from 3.11 of CRLM to 2.54 of Liver-VOI, p<0.001) while uniformity increased (from 0.135 of CRLM to 0.199 of Liver-VOI, p<0.001). At 10 mm from CRLM, entropy was similar to Liver-VOI in 62% of cases and uniformity in 46%. Smaller CRLM and responders to chemotherapy showed higher and earlier normalization of entropy and uniformity values.
Conclusion: The radiomic analysis of peritumoral tissue in CRLM demonstrated a peculiar gradient of decreasing entropy and increasing uniformity despite a normal radiological appearance, representing a potential biomarker for personalized clinical decision-making.
Limitations: Retrospective analysis; small cohort; heterogeneous CT data; missing correlation with pathologic/surgical data.
Funding for this study: AIRC grant #2019−23822
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Protocol 988/22

Author Block: C. Noirot, D. Abler, L. Haefliger, S. Mantziari, M. Schäfer, N. Vietti Violi, A. Depeursinge, C. Dromain, M. Jreige; Lausanne/CH
Purpose: Prognosis evaluation in esophageal cancer remains challenging. Accurate survival prediction is crucial for treatment planning and follow-up strategies. Although MRI and 18F-FDG PET/CT provide valuable information, they have limitations in accurately predicting patient outcomes. This study aimed to develop radiomics models based on MRI and PET/CT to predict overall survival in esophageal cancer patients using baseline and follow-up imaging.
Methods or Background: Sixty patients (M/F: 50/10, mean age 66±9 years) with newly diagnosed esophageal cancer were prospectively included (2017-2022). Patients underwent staging with 18F-FDG PET/CT and MRI, with follow-up MRI after neoadjuvant treatment. Tumors were manually segmented using Mint™ Software, and radiomics features were extracted via QuantImage v2 platform. The dataset, including 645 features from MRI and PET/CT, was split into training (80%) and test (20%) sets. Various survival prediction algorithms were compared. Model performance was assessed with the concordance index (C-index) using bootstrapping for confidence interval (CI) estimation.
Results or Findings: Radiomics features were analyzed at baseline from both PET/CT and MRI for 52 patients, and at follow-up MRI for 49 patients. Mean survival was of 37 months (range: 1.8 to 78.1). The MRI model (14 features) achieved a C-index of 0.733 (95% CI: 0.718–0.756), and the PET/CT model (5 features) achieved 0.724 (95% CI: 0.707–0.746) for predicting OS. A combined model with 19 features improved the C-index to 0.868 (95% CI: 0.853–0.881), while a follow-up MRI model (16 features) reached 0.807 (95% CI: 0.790–0.827).
Conclusion: The radiomics model based on MRI and PET/CT demonstrated robust performance in predicting survival for esophageal cancer patients. Integrating multi-modal baseline and follow-up imaging radiomics features into survival models could enhance prognostic accuracy, improving personalized management strategies in esophageal cancer.
Limitations: The limitations of the study are the number of patients.
Funding for this study: No funding was received.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by "Commission Cantonale d'Ethique de la Recherche sur l'être humain" à Lausanne (CER-VD 2017-00388)

Author Block: K. S. Younis⁷J. Garrett¹, A. Elhanashi², A. Gentili³, S. Faghani⁴, S. Kuanar⁴, Y. Singh⁴, Y. Huo⁵, G. M. Conte⁴, J. Yacoub⁶, O. Unal¹; ¹Madison, WI/US, ²Pisa/IT, ³San Diego, CA/US, ⁴Rochester, MN/US, ⁵Nashville, TN/US, ⁶Washington, DC/US, ⁷Cleveland, OH/US
Purpose: To evaluate the impact of AI-assisted annotation tools on improving consistency in dataset labeling for federated learning, focusing on image segmentation tasks for renal cell carcinoma (RCC) in CT and MR images.
Methods or Background: One of the major challenges in distributed learning is variability in data labels across sites, especially in image segmentation tasks where mask generation methods can differ significantly. This study leverages a publicly available dataset from The Cancer Imaging Archive (TCGA-KIRC) for RCC, which contains rich imaging and clinical data. AI-assisted annotation tools standardize data labeling before training models such as Unet and Swin UNETr. The primary goal is to assess if these tools enhance model performance or reduce the number of cases needed for effective training in federated learning environments. Radiologists from various institutions across multiple states in the US manually annotated the images and federated learning was conducted using NVIDIA nvFLARE.
Results or Findings: Preliminary results suggest that AI-assisted annotation improves model consistency, with segmentation accuracy increasing by approximately 12% when compared to non-standardized data labels. The model efficiency was also reflected in reduced data redundancy and higher annotation agreement between annotators. This enhancement allowed for more precise radiomic analysis across datasets. Additionally, the number of samples required for training decreased by 20%, indicating the efficiency of AI-assisted annotation in generating reliable training datasets.
Conclusion: AI-assisted annotation holds promise for improving performance and efficiency in federated learning environments, particularly for automated image segmentation tasks like RCC detection. The enhanced consistency in label generation helps to reduce the variability introduced by multiple sites, thereby improving model generalizability.
Limitations: The study does not include external validation across diverse imaging platforms, and annotation tools were not evaluated for real-time performance during federated learning.
Funding for this study: Nvidia Education grant
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: N/A