RPS 609 - Hot Topic: AI-driven image interventions

Author Block: D. G. Castiglione, L. Musmeci, G. Failla, F. Libra, D. Falsaperla, F. Vacirca, F. Tiralongo, S. Palmucci, A. Basile; Catania/IT
Purpose: This study aimed to evaluate the capability of two artificial intelligence (AI) models to predict optimal biopsy strategies for pulmonary nodules and to compare their decision-making process with approaches adopted by interventional radiologists.
Methods or Background: We retrospectively reviewed 49 lung biopsy procedures performed at our institution between November 2024 and February 2025. Clinical and imaging data, including lesion size, pleural distance, and complication rates, were collected. Lung nodules were manually delineated on CT scans and submitted in three orthogonal planes to two AI systems. Each AI was asked to simulate key procedural decisions: imaging guidance modality (CT vs US), patient positioning, and needle trajectory. AI outputs were compared with actual biopsy strategies retrieved from RIS/PACS. Subgroup analysis was performed according to operator experience (>10 years vs <10 years). Inter-reader agreement was measured with Cohen’s kappa.
Results or Findings: Overall AI–human agreement was poor (κ = 0.01–0.20). Fair agreement emerged between AI-2 and operators regarding guidance modality and entry point (κ = 0.21–0.40). Moderate agreement was observed when comparing AI-2 with radiologists with >10 years of experience, particularly for selecting CT vs US guidance and trajectory planning (κ = 0.41–0.60). AI-1 displayed weaker overall performance, with minimal alignment to clinical practice.
Conclusion: Current AI tools show limited reliability in replicating expert interventional radiologists’ decision-making in lung biopsy planning. While modest concordance was observed with more experienced operators, both systems lack sufficient robustness for autonomous use. AI may serve as an adjunctive decision-support tool but cannot yet substitute for clinical expertise. Further refinement with larger datasets and advanced training is warranted.
Limitations: N/A
Funding for this study: N/A
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: N/A

Author Block: L. Saccenti, N. Varble, J. Karanian, W. F. Pritchard, M. Li, B. J. Wood; Bethesda, MD/US
Purpose: To evaluate in a phantom and in vivo the accuracy of needle insertions using a smartphone augmented reality (AR) application with integrated needle guide and deep learning-based body tracking registration.
Methods or Background: A smartphone AR application was developed (Unity, Vuforia) with a surface tracking tool based on deep learning, to automatically overlay and scroll CT imaging data on a body, without fiducials or tracking hardware. A smartphone cover with an integrated needle guide (Civco) was designed and 3D printed, permitting use in percutaneous interventions. The target and entry point were selected through the smartphone application. Needle insertion along the planned path was performed with smartphone sensor feedback. Accuracy of needle insertions was assessed in a phantom (Sun Nuclear) (N=30; 5 physicians), and in vivo (N=14 needle insertions, 3 swine) under Institutional Animal Care and Use Committee approval. Accuracy (tip-to-target-center distance) and angular error were evaluated on post-insertion CT (iQon, Philips).
Results or Findings: In phantom, median accuracy was 4.3mm [IQR 2.2-6.7mm], and median angular error was 2.9° [1.5-4.7°]. In vivo, a total of 14 needle insertions (kidney N=5, liver N=4, and tight muscle N=5) were performed. Median accuracy was 8.8mm (IQR 5.5-11.5mm), and median angular error was 3.8° (IQR 2.4-5.7°).
Conclusion: A smartphone augmented reality application enabled automatic overlay of CT with body surface tracking and demonstrated accuracy of 4.3mm in phantom and 8.8mm in vivo for needle insertions. This smartphone AR application has a simple workflow without pre-procedural CT with fiducials or bulky hardware and does not require internal segmentation.
Limitations: Larger studies will include more operators to confirm those results
Funding for this study: Intramural Research grant from the National Institutes of Health
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Institutional Animal Care and Use Committee approval for in vivo study

Author Block: H. Huang, W. Li, W. Li; Shanghai/CN
Purpose: To develop and validate a multi-omics integrated prediction model for complete response (CR) of colorectal cancer (CRC) lung metastases treated by radiofrequency ablation by combining radiomics, pathomics, and clinical features using delta habitat analysis and 2.5D deep learning.
Methods or Background: We analyzed pre- and post-treatment CT images of 249 CRC lung metastases from two centers between August 2016 and June 2022. Delta habitat subregions were generated via K-means clustering of local radiomic features. A 2.5D deep learning approach extracted spatial features from the largest cross-sectional tumor slices, while whole-slide pathology images were processed using weakly supervised multi-instance learning. Feature selection involved the minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) regression. Models were built using support vector machine (SVM), random forest, and other classifiers, and integrated via multi-omics fusion. Performance was evaluated using receiver operating characteristic (ROC) analysis, calibration curves, and decision curve analysis (DCA).
Results or Findings: A total of 150 lung metastases (60%) achieved CR. The Delta Habitat signature achieved an AUC of 0.818 in the validation cohort, outperforming traditional delta radiomics (AUC = 0.741) and pre-habitat models (AUC = 0.740). The 2.5D multi-instance learning and pathomics signatures showed AUCs of 0.711 and 0.780, respectively. The combined multi-omics model achieved the highest AUC of 0.872, with improved calibration and net benefit in DCA.
Conclusion: The integrated multi-omics model significantly enhances CR prediction accuracy by leveraging complementary information from radiomic, pathomic, and clinical data, demonstrating strong potential for supporting clinical decision-making.
Limitations: Larger external validation data is required to confirm generalizability. Computational complexity and dependency on high-quality imaging may limit clinical deployment.
Funding for this study: 1. National Key R&D Program of China (Grant No.2023YFC2411404); 2. Shanghai Anticancer Association EYAS PROJECT (SACA-CY23B03)
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics Committee of Fudan University, Shanghai Cancer Center (Approval Number: 1612167–18).

Author Block: L. Saccenti, O. Pena, L. Hazen, D. Singh, A. Gallagher, A. Mikhail, W. F. Pritchard, J. Karanian, M. Li; Bethesda, MD/US
Purpose: To assess the accuracy of a novel acousto-optic sensor for needle tip localization within ultrasound(US) imaging compared to electromagnetic(EM) tracking in phantoms. The ability to localize tumor and reconstruct US view from the needle tip was also evaluated.
Methods or Background: An acousto-optic 50µm sensor was prototyped and integrated into a 20G stylet. The sensor was interfaced with US and a visualization software (DeepSight Technology). It provided 3D location of the needle tip relative to US plane and an internal US view from the needle tip. A phantom with target tumors (Sun Nuclear) was punctured using the acousto-optic sensor and using a commercial EM tracking system (PercuNav, Philips). The EM tracking system was composed of a US transducer sensor, a needle holder sensor and an EM field generator. Needle insertions were completed when the tracking system indicated the needle tip was located at the center of the target. Accuracies (distances from needle tip to target center) on CT imaging (iQon, Philips) were compared (paired t-test, Rstudio v2024).
Results or Findings: Tumors (5-13cm depth) were targeted using the acousto-optic tracking system (N=10) and then the EM tracking system (N=10). The mean accuracy was 2.85mm (SD 1.32mm) for the acousto-optic tracking system and 3.57mm (SD 1.18mm; P=0.19) for the EM tracking system. Using the acousto-optic system, US imaging from needle tip was feasible for each insertion, with sufficient resolution to allow characterization of target-specific echogenicity around the needle tip. The EM tracking system did not allow internal imaging or tissue characterization capabilities.
Conclusion: This novel acousto-optic sensor could successfully localize the needle tip in 3 dimensions with accuracy similar to EM tracking in phantom, while enabling internal US imaging from the needle tip.
Limitations: Phantom study
Funding for this study: NIH intramural grant
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: M. M. Jabeer, S. Tyagi, J. Singh, A. Chandalia; Bengaluru/IN
Purpose: Pancreatic cancer remains one of the deadliest malignancies, with approximately 510,922 new cases and 467,409 deaths reported globally in 2022. Although it ranks as the 12th most common cancer, it is the sixth leading cause of cancer-related mortality. The poor prognosis is primarily due to late diagnosis and a lack of effective early detection strategies. Accurate segmentation of pancreatic tumors in abdominal CT imaging is essential but labor-intensive for diagnosis, treatment planning, and surgical guidance. This study aims to develop and evaluate a deep learning model for automated segmentation and reporting of pancreatic tumors on abdominal CT scans .
Methods or Background: We designed a deep learning pipeline architecture to segment the pancreas and pancreatic lesions. The model was trained on a retrospective dataset of 2,000 abdominal CT scans, including both healthy individuals and patients with pancreatic cancer. All scans were manually annotated by experienced radiologists, providing high-quality ground truths. The two-stage pipeline first delineates the pancreas, followed by lesion segmentation within the extracted region. The output includes a DICOM-compatible overlay and an automatically generated report detailing lesion count, size, and volume.
Results or Findings: On a test set of 100 abdominal CT cases, the model achieved a mean Dice coefficient of 0.86 ± 0.09, sensitivity of 93.6%, and specificity of 88.4%. The segmentation results were consistent across different tumor sizes and anatomical variations and were validated by a radiologist for clinical reliability.
Conclusion: Our study demonstrates the feasibility of an automated deep learning approach for pancreatic tumor segmentation on CT scans. The model achieved high accuracy, offering potential to enhance diagnosis, treatment planning, and clinical workflows.
Limitations: Multi-center, prospective validation on more diverse datasets will be essential to confirm robustness across imaging protocols and clinical settings.
Funding for this study: No funding was obtained for this work.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: S. De la Chica Bolaños, J. Salazar, D. C. Gimenez, A. Soldevila, F. A. Armas Terry, S. Guillén Rodriguez, N. De La Torre Rubio, R. Castañeda Argaiz, M. Figols Gorina; Manresa/ES
Purpose: To describe our experience with robotic in-bore MRI-guided prostate biopsy with the Soteria® system in PI-RADS 4 and 5 lesions within a fast-track diagnostic pathway at a public university hospital, and to analyse effectiveness and safety.
Methods or Background: A single-centre retrospective study (01.08.2024–31.07.2025) was conducted at Hospital Sant Joan de Déu – Fundació Althaia, Manresa, Spain, in patients referred from a fast-track diagnostic pathway and undergoing in-bore MRI-guided prostate biopsy with the Soteria® system for PI-RADS 4 or 5 lesions. Clinical data, MRI lesion characteristics (size, location), and histopathological outcomes were analysed. Complications (rectal bleeding, haematuria, fever, emergency visits) were recorded.
Results or Findings: 72 cases were analysed. Mean age was 71 ± 8.1 years and median PSA 8.8 ng/mL (IQR 5.7–14.6). The malignancy rate was 61.1% (95% CI: 48.9–72.4). Of these, 47 were PI-RADS 4 and 24 PI-RADS 5. Malignancy was more frequent in PI-RADS 5 than in PI-RADS 4 (87.5% vs 48.9%; p=0.0015). Most tumours were in the peripheral zone (68.2%), followed by transitional (27.3%) and central (2.3%). Histologically, PI-RADS 5 lesions showed a higher proportion of ISUP grade ≥4 compared with PI-RADS 4 lesions (47.6%; p=0.033). Concordance between biopsy and surgical specimen ISUP was substantial (Weighted Kappa=0.77, 95% CI: 0.49–1.00) in the five patients with surgery. Complications were infrequent, reported in three of 54 patients with follow-up (5.6%): one mild rectal bleeding, two febrile episodes, and one emergency visit.
Conclusion: Robotic in-bore MRI-guided prostate biopsy is safe, effective, and achieves high cancer detection. PI-RADS 5 correlated with malignancy and higher ISUP. Substantial concordance was observed between biopsy and surgical ISUP.
Limitations: The small number of surgical cases for ISUP comparison.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: A. Posa, P. Barbieri, M. Lippi, G. D'Aniello, E. V. Andreani, R. Iezzi; Rome/IT
Purpose: To evaluate the effectiveness of AI-driven predictive models in predicting HCC recurrence.
Methods or Background: Recurrence prediction of hepatocellular carcinoma (HCC) after thermal ablation represents a challenge that can impact patients' quality of life. Artificial Intelligence (AI)-based radiomics models, applied to various imaging modalities, can improve recurrence prediction, therefore guiding therapeutic decisions.
A systematic literature search in PubMed and Scopus has been performed. A total of 17 studies were selected to be included in this systematic review. All studies included response prediction evaluation with AI models for patients who underwent thermal ablation for HCC. Deep learning and machine learning algorithms were confronted to evaluate the predictive performance and accuracy using metrics such as the area under the curve (AUC) and concordance index (C-index).
Results or Findings: The developed models demonstrated high accuracy in predicting local progression and recurrence, allowing a solid risk stratification. In particular, the integration of imaging data and clinical-laboratory variables optimized treatment selection, highlighting the superior ability of imaging models to predict therapeutic outcomes compared to clinical parameters alone. Furthermore, radiomic analysis of follow-up imaging enabled highly accurate detection of ablation site recurrence.
Conclusion: AI-driven predictive models based on multimodal radiomic analyses integrated with clinical data represent promising tools for predicting tumour recurrence after thermal ablation in HCC patients.
Limitations: Mostly retrospective and monocentric studies. Different reported outcome measures.
Funding for this study: None.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: C. KONG, S. Zheng, Y. Su; Lishui/CN
Purpose: To develop and validate a multimodal model integrating preoperative MR radiomics and postoperative H&E pathomics to predict recurrence after hepatectomy with adjuvant transarterial chemoembolization (TACE) in hepatocellular carcinoma (HCC).
Methods or Background: We retrospectively included 102 HCC patients who underwent curative resection and received adjuvant TACE within 2 months. On preoperative MR-T2WI, tumor and 1/3/5-mm peritumoral rings were segmented to extract 9,056 radiomic features. From postoperative H&E whole-slide images, 1,079 pathomic features were derived. Feature selection used maximum relevance–minimum redundancy (mRMR) and least absolute shrinkage and selection operator (LASSO)–Cox, retaining 5 radiomic and 9 pathomic features, from which logistic regression derived Radiomics score (RadScore) and Pathomics score (PathScore). Patients were stratified by each biomarker, and Kaplan–Meier analyses compared recurrence-free survival (RFS) between biomarker-defined strata. Prespecified clinical covariates (e.g., BCLC stage, antiviral therapy) were tested by univariable Cox, with significant/a-priori factors entered into a multivariable model. Seven prediction models were trained: imaging-only, pathology-only, clinical-only, imaging+clinical, pathology+clinical, imaging+pathology, and fusion model (imaging+pathology+clinical). Model performance was evaluated in training and validation cohorts using time-dependent receiver operating characteristic (ROC) curves and area under the curve (AUC) at 1, 2, and 3 years, calibration, and decision-curve analysis (DCA).
Results or Findings: RadScore- and PathScore-defined groups showed significantly different RFS (log-rank p<0.05). Antiviral therapy emerged as an independent prognostic factor. The fusion model achieved strong discrimination with training AUCs of 0.836/0.859/0.932 and validation AUCs of 0.807/0.823/0.814 at 1/2/3 years. Calibration indicated agreement between predicted and observed RFS, and DCA demonstrated favorable net clinical benefit across relevant thresholds.
Conclusion: Integrating peri-tumoral MR radiomics with postoperative H&E pathomics, complemented by key clinical variables, enables accurate prediction of recurrence after adjuvant TACE in resected HCC and supports individualized post-operative risk stratification and surveillance.
Limitations: Single center study.
Funding for this study: None.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The Institutional Review Board and Human Ethics Committee of the Fifth Afliated Hospital of Wenzhou Medical University.