Effective radiotherapy planning requires precise delineation of organs at risk (OARs), but the traditional manual method is laborious and subject to variability. This study explores using convolutional neural networks (CNNs) for automating OAR segmentation in pelvic CT images, focusing on the bladder, prostate, rectum, and femoral heads (FHs) as an efficient alternative to manual segmentation. Utilizing the Medical Open Network for AI (MONAI) framework, we implemented and compared U-Net, ResU-Net, SegResNet, and Attention U-Net models and explored different loss functions to enhance segmentation accuracy. Our study involved 240 patients for prostate segmentation and 220 patients for the other organs. The models' performance was evaluated using metrics such as the Dice similarity coefficient (DSC), Jaccard index (JI), and the 95th percentile Hausdorff distance (95thHD), benchmarking the results against expert segmentation masks. SegResNet outperformed all models, achieving DSC values of 0.951 for the bladder, 0.829 for the prostate, 0.860 for the rectum, 0.979 for the left FH, and 0.985 for the right FH (p < 0.05 vs. U-Net and ResU-Net). Attention U-Net also excelled, particularly for bladder and rectum segmentation. Experiments with loss functions on SegResNet showed that Dice loss consistently delivered optimal or equivalent performance across OARs, while DiceCE slightly enhanced prostate segmentation (DSC = 0.845, p = 0.0138). These results indicate that advanced CNNs, especially SegResNet, paired with optimized loss functions, provide a reliable, efficient alternative to manual methods, promising improved precision in radiotherapy planning.

To investigate vessel delineation and image quality of ultra-high-resolution (UHR) CT angiography (CTA) reconstructed using deep learning reconstruction (DLR) optimised for brain CTA (DLR-brain) in moyamoya disease (MMD), compared with DLR optimised for body CT (DLR-body) and hybrid iterative reconstruction (Hybrid-IR). This retrospective study included 50 patients with suspected or diagnosed MMD who underwent UHR brain CTA. All images were reconstructed using DLR-brain, DLR-body, and Hybrid-IR. Quantitative analysis focussed on moyamoya perforator vessels in the basal ganglia and periventricular anastomosis. For these small vessels, edge sharpness, peak CT number, vessel contrast, full width at half maximum (FWHM), and image noise were measured and compared. Qualitative analysis was performed by visual assessment to compare vessel delineation and image quality. DLR-brain significantly improved edge sharpness, peak CT number, vessel contrast, and FWHM, and significantly reduced image noise compared with DLR-body and Hybrid-IR (P < 0.05). DLR-brain significantly outperformed the other algorithms in the visual assessment (P < 0.001). DLR-brain provided superior visualisation of small intracranial vessels compared with DLR-body and Hybrid-IR in UHR brain CTA.

C2 pars interarticularis length (C2PIL) required for pars screws has not been thoroughly studied in subjects with high-riding vertebral artery (HRVA). We aimed to measure C2PIL specifically on the sides with HRVA, define short pars, optimal pars screw length, and incorporate C2PIL into HRVA clusters using machine learning algorithms. A clinical anatomical study based on cervical CT was conducted with STROBE-compliant case-control design. HRVA was defined as accepted. Interobserver, intraobserver, and inter-software agreement coefficients for HRVA were adopted from our previous study. Sample size was estimated with pwr package and C2PIL was measured. Cut-off value and predictive statistics of C2PIL for HRVA were computed with cutpointr package. Unsupervised machine learning clustering was applied with all three pars parameters. 345 potential screw insertion sites (PSIS) were grouped as HRVA (143 PSIS in 110 subjects) or controls (202 PSIS in 101 subjects). 68% participants were females. The median C2PIL in HRVA group was 13.7 mm with interquartile range (IQR) of 1.7, whereas in controls it was 19.8 mm (IQR = 2.7). The optimal cut-off value of C2PIL discriminating HRVA was 16.06 mm with sensitivity of 96.5% and specificity of 99.3%. Therefore, clinically important short pars was defined as ≤ 16 mm rounding to the nearest screw length. Two clusters were created incorportating three parameters of pars interarticularis. In preoperative planning, the identified C2PIL cut-off of ≤ 16 mm may assist surgeons in early recognition of HRVA. The average screw lengths of 14 mm for bicortical and 12 mm for safer unicortical purchase in HRVA cases may serve as practical intraoperative reference points, particularly in situations requiring rapid decision-making or when navigation systems are unavailable. Moreover, C2PIL complements the classic HRVA parameters within the dichotomized clustering framework.

Lung cancer is the leading cause of cancer mortality worldwide, and non-invasive methods for detecting key mutations and staging are essential for improving patient outcomes. Here, we compare the performance of two machine learning models - FMCIB+XGBoost, a supervised model with domain-specific pretraining, and Dinov2+ABMIL, a self-supervised model with attention-based multiple-instance learning - on 3D lung nodule data from the Stanford Radiogenomics and Lung-CT-PT-Dx cohorts. In the task of KRAS and EGFR mutation detection, FMCIB+XGBoost consistently outperformed Dinov2+ABMIL, achieving accuracies of 0.846 and 0.883 for KRAS and EGFR mutations, respectively. In cancer staging, Dinov2+ABMIL demonstrated competitive generalization, achieving an accuracy of 0.797 for T-stage prediction in the Lung-CT-PT-Dx cohort, suggesting SSL's adaptability across diverse datasets. Our results emphasize the clinical utility of supervised models in mutation detection and highlight the potential of SSL to improve staging generalization, while identifying areas for enhancement in mutation sensitivity.

This work aimed to develop an automated method for quantifying the distribution and severity of perfusion changes on CT pulmonary angiography (CTPA) in patients with chronic thromboembolic pulmonary hypertension (CTEPH) and to assess their associations with clinical parameters and expert annotations. Following automated segmentation of the chest, a machine-learning model assuming three distributions of attenuation in the pulmonary parenchyma (hyperemic, normal, and oligemic) was fitted to the attenuation histogram of CTPA images using Bayesian analysis. The proportion of each component, its spatial heterogeneity (entropy), and center-to-periphery distribution of the attenuation were calculated and correlated with the findings on CTPA semi-quantitatively evaluated by radiologists and with clinical function tests. CTPA scans from 52 patients (mean age, 65.2 ± 13.0 years; 27 men) diagnosed with CTEPH were analyzed. An inverse correlation was observed between the proportion of normal parenchyma and brain natriuretic propeptide (proBNP, ρ = -0.485, p = 0.001), mean pulmonary arterial pressure (ρ = -0.417, p = 0.002) and pulmonary vascular resistance (ρ = -0.556, p < 0.0001), mosaic attenuation (ρ = -0.527, p < 0.0001), perfusion centralization (ρ = -0.489, p = < 0.0001), and right ventricular diameter (ρ = -0.451, p = 0.001). The entropy of hyperemic parenchyma showed a positive correlation with the pulmonary wedge pressure (ρ = 0.402, p = 0.003). The slope of center-to-periphery attenuation distribution correlated with centralization (ρ = -0.477, p < 0.0001), and with proBNP (ρ = -0.463, p = 0.002). This study validates an automated system that leverages Bayesian analysis to quantify the severity and distribution of perfusion changes in CTPA. The results show the potential of this method to support clinical evaluations of CTEPH by providing reproducible and objective measures. Question This study introduces an automated method for quantifying the extent and spatial distribution of pulmonary perfusion abnormalities in CTEPH using variational Bayesian estimation. Findings Quantitative measures describing the extent, heterogeneity, and distribution of perfusion changes demonstrate strong correlations with key clinical hemodynamic indicators. Clinical relevance The automated quantification of perfusion changes aligns closely with radiologists' evaluations, delivering a standardized, reproducible measure with clinical relevance.

To assess the predictive performance, risk stratification capabilities, and auxiliary diagnostic utility of radiomics, deep learning, and fusion models in identifying visceral pleural invasion (VPI) in lung adenocarcinoma. A total of 449 patients (female:male, 263:186; 59.8 ± 10.5 years) diagnosed with clinical IA stage lung adenocarcinoma (LAC) from two distinct hospitals were enrolled in the study and divided into a training cohort (n = 289) and an external test cohort (n = 160). The fusion models were constructed from the feature level and the decision level respectively. A comprehensive analysis was conducted to assess the prediction ability and prognostic value of radiomics, deep learning, and fusion models. The diagnostic performance of radiologists of varying seniority with and without the assistance of the optimal model was compared. The late fusion model demonstrated superior diagnostic performance (AUC = 0.812) compared to clinical (AUC = 0.650), radiomics (AUC = 0.710), deep learning (AUC = 0.770), and the early fusion models (AUC = 0.586) in the external test cohort. The multivariate Cox regression analysis showed that the VPI status predicted by the late fusion model were independently associated with patient disease-free survival (DFS) (p = 0.044). Furthermore, model assistance significantly improved radiologist performance, particularly for junior radiologists; the AUC increased by 0.133 (p < 0.001) reaching levels comparable to the senior radiologist without model assistance (AUC: 0.745 vs. 0.730, p = 0.790). The proposed decision-level (late fusion) model significantly reducing the risk of overfitting and demonstrating excellent robustness in multicenter external validation, which can predict VPI status in LAC, aid in prognostic stratification, and assist radiologists in achieving higher diagnostic performance.

Deformable image registration (DIR) is an enabling technology in many diagnostic and therapeutic tasks. Despite this, DIR algorithms have limited clinical use, largely due to a lack of benchmark datasets for quality assurance during development. DIRs of intra-patient abdominal CTs are among the most challenging registration scenarios due to significant organ deformations and inconsistent image content. To support future algorithm development, here we introduce our first-of-its-kind abdominal CT DIR benchmark dataset, comprising large numbers of highly accurate landmark pairs on matching blood vessel bifurcations. Abdominal CT image pairs of 30 patients were acquired from several publicly available repositories as well as the authors' institution with IRB approval. The two CTs of each pair were originally acquired for the same patient but on different days. An image processing workflow was developed and applied to each CT image pair: (1) Abdominal organs were segmented with a deep learning model, and image intensity within organ masks was overwritten. (2) Matching image patches were manually identified between two CTs of each image pair. (3) Vessel bifurcation landmarks were labeled on one image of each image patch pair. (4) Image patches were deformably registered, and landmarks were projected onto the second image. (5) Landmark pair locations were refined manually or with an automated process. This workflow resulted in 1895 total landmark pairs, or 63 per case on average. Estimates of the landmark pair accuracy using digital phantoms were 0.7 mm ± 1.2 mm. The data are published in Zenodo at https://doi.org/10.5281/zenodo.14362785. Instructions for use can be found at https://github.com/deshanyang/Abdominal-DIR-QA. This dataset is a first-of-its-kind for abdominal DIR validation. The number, accuracy, and distribution of landmark pairs will allow for robust validation of DIR algorithms with precision beyond what is currently available.

The aim of this study was to retrain our existing deep learning-based fully automated aortic valve leaflets/root measurement algorithm, using computed tomography (CT) data for root dilatation (RD), and assess its clinical feasibility. 67 ECG-gated cardiac CT scans were retrospectively collected from 40 patients with RD to retrain the algorithm. An additional 100 patients' CT data with aortic stenosis (AS, n=50) and aortic regurgitation (AR) with/without RD (n=50) were collected to evaluate the algorithm. 45 AR patients had RD. The algorithm provided patient-specific 3-dimensional aortic valve/root visualization. The measurements of 100 cases automatically obtained by the algorithm were compared with an expert's manual measurements. Overall, there was a moderate-to-high correlation, with differences of 6.1-13.4 mm²for the virtual basal ring area, 1.1-2.6 mm for sinus diameter, 0.1-0.6 mm for coronary artery height, 0.2-0.5 mm for geometric height, and 0.9 mm for effective height, except for the sinotubular junction of the AR cases (10.3 mm) with an indefinite borderline over the dilated sinuses, compared with 2.1 mm in AS cases. The measurement time (122 s) per case by the algorithm was significantly shorter than those of the experts (618-1,126 s). This fully automated algorithm can assist in evaluating aortic valve/root anatomy for planning surgical and transcatheter treatments while saving time and minimizing workload.

This study aimed to compare two deep learning reconstruction (DLR) techniques (AiCE mild; AiCE strong) with two established methods-iterative reconstruction (IR) and filtered back projection (FBP)-for the detection of monosodium urate (MSU) in dual-energy computed tomography (DECT). An ex vivo bio-phantom and a raster phantom were prepared by inserting syringes containing different MSU concentrations and scanned in a 320-rows volume DECT scanner at different tube currents. The scans were reconstructed in a soft tissue kernel using the four reconstruction techniques mentioned above, followed by quantitative assessment of MSU volumes and image quality parameters, i.e., signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR). Both DLR techniques outperformed conventional IR and FBP in terms of volume detection and image quality. Notably, unlike IR and FBP, the two DLR methods showed no positive correlation of the MSU detection rate with the CT dose index (CTDIvol) in the bio-phantom. Our study highlights the potential of DLR for DECT imaging in gout, where it offers enhanced detection sensitivity, improved image contrast, reduced image noise, and lower radiation exposure. Further research is needed to assess the clinical reliability of this approach.

This is a multi-institutional study to evaluate a head-and-neck CT auto-segmentation software across seven institutions globally. 11 lymph node levels and 7 organs-at-risk contours were evaluated in a two-phase study design. Time savings were measured in both phases, and the inter-observer variability across the seven institutions was quantified in phase two. Overall time savings were found to be 42% in phase one and 49% in phase two. Lymph node levels IA, IB, III, IVA, and IVB showed no significant time savings, with some centers reporting longer editing times than manual delineation. All the edited ROIs showed reduced inter-observer variability compared to manual segmentation. Our study shows that auto-segmentation plays a crucial role in harmonizing contouring practices globally. However, the clinical benefits of auto-segmentation software vary significantly across ROIs and between clinics. To maximize its potential, institution-specific commissioning is required to optimize the clinical benefits.