RPS 1912 - Hot Topic: AI in paediatric radiology

Author Block: K. Saravanan¹, F. Abubacker Sulaiman², R. Praveenkumar², J. Lydia², D. Velan²; ¹Melmaruvathur/IN, ²Chennai/IN
Purpose: To develop and validate an artificial intelligence (AI)–based radiomics model integrating diffusion-weighted imaging (DWI) and T2-weighted MRI features of the placenta for early prediction of preeclampsia (PE) and fetal growth restriction (FGR). The objective is to identify microstructural and textural biomarkers preceding clinical manifestation, enabling proactive obstetric management and improved perinatal outcomes.
Methods or Background: This prospective study included 80 pregnant women (20–32 weeks gestation) who underwent placental MRI on a 3-Tesla scanner. DWI (b-values 0, 800 s/mm²) and high-resolution T2 sequences were analyzed. Placental volumes were segmented semi-automatically, and radiomic features (first-order, texture, and shape) were extracted. Machine learning models—including Random Forest and Gradient Boosting—were trained to classify risk for PE/FGR, with cross-validation and feature selection via LASSO regression.

Clinical parameters (blood pressure, uterine artery Doppler indices) were integrated into a multimodal model for comparison.
Results or Findings: The AI-radiomics model demonstrated an AUC of 0.93 for predicting PE and 0.91 for FGR, outperforming conventional Doppler parameters (AUC 0.76). Key discriminative features included gray-level co-occurrence entropy and ADC histogram kurtosis, reflecting placental microstructural heterogeneity. Combined clinical-radiomics integration improved specificity and early detection (mean 5 weeks before clinical onset).
Conclusion: AI-enhanced placental radiomics from DWI and T2 MRI enables noninvasive early prediction of preeclampsia and FGR with high accuracy. This approach holds promise for precision obstetric imaging, facilitating timely intervention and reducing perinatal morbidity.
Limitations: Single-center design, limited sample size, and variability in placental segmentation may restrict generalizability. Lack of histopathological correlation and absence of external validation cohorts warrant further multicentric studies for clinical translation.
Funding for this study: Not Applicable
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: B. Wichtlhuber¹, E. Frodl¹, M. Sayed¹, T. Persigehl², M. Neitzel¹, J. Dietz¹, M. Eicke³, D. M. Renz³, A. M. Bucher¹; ¹Frankfurt/DE, ²Koeln/DE, ³Hannover/DE
Purpose: Paediatric lymph node segmentation faces critical data scarcity challenges. We hypothesised that incorporating adult CT data into deep learning training would overcome this limitation while maintaining clinical relevance for paediatric oncology applications, particularly for detecting pathologically enlarged nodes requiring follow-up.
Methods or Background: We analysed 418 head/neck CTs from University Hospital Frankfurt: 146 paediatric (<18 years) and 272 adult cases. A 3D full-resolution nnU-Net underwent five-fold cross-validation training (250 epochs) using three strategies: paediatric-only, adult-only, and combined training. Approximately 20% of cases were reserved for independent testing. Evaluation metrics included the Dice coefficient, Intersection over Union (IoU). Clinical relevance was assessed through volumetric coverage analysis, with particular focus on nodes >10mm short-axis diameter—the threshold for pathological enlargement requiring clinical action. Values are presented as median and standard deviation.
Results or Findings: Cross-age training outperformed paediatric-only approaches. The combined model achieved Dice=0.714 ±0.151 and IoU=0.575 ±0.176 on paediatric test data versus Dice=0.695 ±0.155 and IoU=0.553 ±0.179 for paediatric-only training, showing clear improvement. Interestingly, this benefit was unidirectional—paediatric patients gained from mixed training, whereas adult test performance remained unchanged (combined model Dice=0.643 ±0.125 vs. adult-only Dice=0.647 ±0.113). Clinical utility analysis showed strong performance: 88.97% of enlarged nodes (>10mm) achieved ≥10% volumetric coverage, ensuring reliable oncological detection. Moreover, 69.08% of all lymph nodes reached ≥50% coverage, supporting accurate volumetric measurements. These metrics translate into improved workflow efficiency, reducing missed findings while preserving precision for follow-up.
Conclusion: Cross-age training addresses the key challenge of paediatric data scarcity in medical imaging AI, improving technical metrics and enabling clinically meaningful detection of pathological lymphadenopathy. This approach supports robust AI use in paediatric radiology with limited annotated data, potentially accelerating adoption in oncology workflows.
Limitations: A limitation of the study is its single-center validation.
Funding for this study: Funding was provided by the German Federal Ministry of Education and Research through the RACOON project (reference number 01KX2021).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics approval was obtained by University Medicine Frankfurt (Reference 2023-1459)

Author Block: L. Ibrahim, A. Ilivitzki, M. Freiman; Haifa/IL
Purpose: To investigate whether radiomic features from post–first-cycle PET and CT scans, alone or combined with clinical variables, improve the prediction of event-free survival (EFS) in pediatric high-risk Hodgkin lymphoma (HL, stage IIIB/IVB), compared with a baseline model incorporating clinical variables and radiological PET assessment.
Methods or Background: Outcome prediction in pediatric HL traditionally relies on risk-adapted protocols combining clinical factors with interim FDG PET/CT, mainly focused on the isotopic PET component. Prior evidence suggests that CT provides additional prognostic information, underscoring the importance of anatomical features.
In this study, post–first-cycle PET/CT scans from 137 patients enrolled in the Children’s Oncology Group AHOD0831 trial (NCT01026220) were analyzed. Up to five lesions per patient were segmented; PET images were SUV-standardized, and radiomic features from PET and CT were extracted and averaged using tumor volume weighting. Clinical variables included demographics, histology, stage, bulky disease, and radiological PET response. Machine learning models were trained with stratified 5-fold cross-validation, and performance metrics were averaged across folds
Results or Findings: EFS events occurred in 27 patients (19.7%). CT radiomics outperformed the baseline clinical + PET model (accuracy: 0.78 ± 0.10; balanced accuracy: 0.71 ± 0.09 vs. 0.70 ± 0.11; 0.66 ± 0.04) and exceeded PET radiomics (accuracy: 0.77 ± 0.07; balanced accuracy: 0.66 ± 0.07). The strongest predictive performance was achieved when CT radiomics were combined with clinical variables.
Conclusion: Radiomics from CT improved prognostic prediction beyond the standard clinical + PET model and outperformed PET radiomics, reinforcing the anatomical contribution to risk stratification. CT radiomics represent a promising complementary tool for early prognostic assessment in pediatric high-risk HL.
Limitations: This study was based on internal cross-validation within a single trial cohort (AHOD0831), without external or temporal validation, which may limit generalizability.
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information:

Author Block: I. Bianco, D. Ippolito, C. Maino, C. R. G. L. O. M. Talei Franzesi, P. N. Franco, D. G. Gandola, R. Corso; Monza/IT
Purpose: To compare image quality and radiation dose between deep learning reconstruction (DLR) and hybrid iterative reconstruction (HIR) algorithms in unenhanced chest CT scans of pediatric patients.
Methods or Background: We retrospectively reviewed 142 unenhanced, single-phase chest CT scans performed for routine diagnostic purpose. 71 pediatric patients were examined using a 128-slice MDCT scanner (100kV) with a DLR algorithm (Precise Image), while the control group of 71 patients underwent scanning with a 256-slice MDCT scanner (100kV) using HIR algorithm (iDose4). Subjective image quality was assessed using a 5-point Likert scale, and objective quality was evaluated by measuring Hounsfield Units (HU) and Standard Deviations (SD) in lung parenchyma, tracheal lumen, air, aorta and muscle. Radiation dose metrics (CTDIvol and DLP) were recorded for both groups, and the estimated effective dose (EED) was calculated using age-specific k-factors.
Results or Findings: A total of 142 pediatric patients were included (median age:10 years).Inter-reader agreement for Likert scale image quality assessment was moderate (κ=0.432) with significantly higher subjective image quality scores for DLR group compared to HIR group (p=0.03).Quantitatively, air and muscle HU differed significantly between the two scanners (air, p=0.036; muscle, p<0.001). Image noise, assessed by SD was significantly lower in DLR group across the lung parenchyma (p=0.007), tracheal lumen (p=0.002) and muscle (p<0.001). Radiation dose metrics were significantly reduced in DLR group compared to control group (mean DLP: 49.6 vs 188.9 mGy·cm;CTDIvol: 1.6 vs 5.6 mGy;EED:0.8 vs 3.2 mSv;all p<0.001).Age-based sub-analysis using quartiles (0–5,6–12,>13 years) confirmed a consistent threefold reduction in DLP and CTDIvol, and a fourfold reduction in EED across all age groups (all p<0.001).
Conclusion: DLR in unenhanced chest CT allows significant radiation dose reduction in pediatric population while delivering superior image quality compared to HIR algorithms.
Limitations: Monocentric study
Funding for this study: None
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: