Abstract Background Despite concerns about radiation dose in pediatric patients, computed tomography (CT) remains a first‐line imaging modality for children. Spectral CT provides additional diagnostic information that benefits pediatric imaging. However, most existing spectral CT systems are primarily optimized for adult patients. Purpose To develop and evaluate a novel double bowtie filter integrating a K‐edge material layer with a conventional Teflon filter for CT. The proposed design aims to enhance spectral signal‐to‐noise ratio (SNR) and spectral separation within the projection domain while maintaining radiation dose levels suitable for pediatric imaging. Methods A simulation framework was developed and used to model a rapid kVp‐switching CT system operating at 70/110 kVp with realistic tube power and geometric constraints. Pediatric phantoms of three sizes (100–200 mm anterior–posterior width) were used to evaluate performance. Five accessible and safe filter materials—gadolinium (Gd), holmium (Ho), erbium (Er), silver (Ag), and tin (Sn)—were tested in combination with a Teflon bowtie. System performance was quantified using virtual monoenergetic image (VMI) SNR at 40 and 70 keV, and the area under the monoenergetic SNR curve (AUMC) as a comprehensive spectral image quality metric. Dose consistency was ensured using a traditional Teflon bowtie as the reference. Results The Teflon + Gd configuration achieved the highest performance, improving mean AUMC over all fan angles by 53.6% on average and up to 56% for the largest phantom. VMI SNR increased by approximately 46% at 40 keV and 42% at 70 keV. Conclusions The double‐bowtie concept substantially enhances spectral performance. The Teflon + Gd design provides a manufacturable, pediatric‐optimized solution adaptable to kVp‐switching and other spectral CT architectures, offering improved diagnostic quality at low dose levels.
Abstract Background Deep learning algorithms for image processing have been proposed for a wide range of applications. In particular, deep learning CT denoising has demonstrated superior ability to reduce noise compared to conventional methods. However, the nonlinear nature of these algorithms makes it difficult to inspect and visualize how the model accomplishes the denoising task. Purpose We propose a local linearization approach to characterize and analyze the behavior of deep learning algorithms. Example analysis was performed for a classic denoising architecture with convolutional neural network and ReLU activation functions, and was extended to GELU activation and a transformer‐based architecture. Methods We defined strictly local linear regions as the input space over which the activation patterns of all ReLU activation functions are the same. The extent of these regions was investigated by perturbing clinically relevant features of a lung lesion. We further defined approximately local linear regions as the input space where Jacobians are similar and linear approximation results in small error. We developed an algorithm to partition input space over a range of lesion and background features into approximately local linear regions, and separately quantified the contribution of the noise realization and anatomical background to this approximation error. The behavior of networks trained with different levels of denoising, as well as alternative architectures, was compared. Results The extent of strictly local linear regions was found to be very small. Perturbations on the scale of were enough to push the network to a neighboring strictly local linear region. However, the Jacobians over a large number of strictly local linear regions were found to be highly similar, which suggests that the network behaves approximately as a piecewise linear system. The partitioning algorithm identified a small number of approximately local linear regions within input spaces characterized by clinically relevant features; this number was similar for the GELU variant but increased for networks trained with more aggressive noise reduction and for a transformer‐based architecture, while Jacobians transferred across anatomical backgrounds also increased approximation error. Conclusions The local linearity analysis framework can be a useful tool for elucidating the behavior of neural networks that are traditionally considered as “black‐boxes.” Preliminary results suggest that the framework can be applied to other differentiable architectures, subject to empirical validation, to characterize these highly nonlinear, data‐dependent algorithms over input spaces of interest.
Abstract Background In two‐dimensional transthoracic echocardiography (TTE), the standard parasternal long‐axis (PLAX) view enables reliable evaluation of the aortic valve and aortic root. Owing to the limited field of view of the ultrasound probe and rib‐window occlusion, the distal ascending aorta (AA) is frequently truncated, compromising the diagnosis of aortic aneurysm and dilation. Purpose We aim to address the incomplete visualization of the distal AA in standard PLAX views. By aggregating fragmented anatomical features intermittently visible in adjacent frames, we enhance the clarity of vessel wall edges and generate the distal AA structure, thereby completing the missing AA imagery in standard PLAX views. Materials and methods We retrospectively collected 1160 TTE videos from Nanjing Drum Tower Hospital (August 2022 – December 2024), all showing complete AA anatomy. After de‐identification, 1059 cases were randomly split into training/validation/test sets (6:2:2), and 101 cases were reserved for observer agreement experiments. An additional 152 videos from three different device models (EPIQ CVX, GE Voluson E10, Resona A20T) formed an external validation set. Two training data types were constructed: (i) dual‐view synchronous data – a standard PLAX sequence and an AO‐PLAX sequence acquired after a cranial probe shift, phase‐matched at end diastole (ED); and (ii) locally masked data – the distal AA region was masked, forcing the model to infer missing anatomy from left‐side context. Input sequences of five frames (ED 2) were extracted. The proposed Spatio‐Temporal Ascending‐Aorta generative adversarial network (STAA‐GAN) first aligns adjacent frames using an Edge Boosted Feature Alignment (EBFA) module with multi‐stage modulated deformable convolutions. A channel‐wise fusion strategy then aggregates temporally aligned features. The fused representation is fed into a Swin‐UNet generator enhanced with a Strip based Cross window Prediction (SCP) module to capture long‐range structural dependencies. Training employs a composite loss including pixel reconstruction, feature reconstruction, texture consistency, relativistic average least‐squares GAN (RaLSGAN), temporal coherence, and structure‐aware terms, with data‐dependent weighting. Statistical analyzes included Bland‐Altman, Passing‐Bablok regression, paired t ‐tests with Cohen's d, Bonferroni correction for multiple comparisons, and two one‐sided tests (TOST, ). Results STAA‐GAN achieved the best image‐quality values among the compared methods, with peak signal‐to‐noise ratio (PSNR) = 22.51 dB, structural similarity index (SSIM) = 0.7574, multi‐scale structural similarity index (MS‐SSIM) = 0.8598, and learned perceptual image patch similarity (LPIPS) = 0.1758. All 12 pairwise comparisons (SRN, IOH, and U‐Transformer versus STAA‐GAN, designated as the reference group) remained statistically significant after Bonferroni correction (), with absolute effect sizes ranging from 0.255 to 3.439. The per‐frame inference time was 203 ms. In clinical consistency validation, the mean difference between measurements on generated and original images was cm with 95% limits of agreement within cm; TOST with cm confirmed clinical equivalence for all comparisons (). The segmentation Dice coefficient on generated images averaged across five cross‐validation folds. External validation on three different ultrasound devices (Philips, GE, Mindray) maintained consistent equivalence (TOST , Cohen's ). These results demonstrate that STAA‐GAN produces clinically interchangeable distal ascending aorta images with high fidelity and real‐time feasibility. Conclusions STAA‐GAN enables extrapolation of the missing distal ascending aorta in standard PLAX views, achieving improved image quality and clinically equivalent diameter measurements on both internal and external validation datasets. These findings support its potential for assisting distal ascending aorta assessment in routine echocardiography.
Abstract Purpose Magnetic particle imaging (MPI) is an emerging biomedical imaging modality. In recent years, human‐scale MPI systems based on field‐free line MPI (FFL‐MPI) projection imaging have been increasingly advanced. However, the lack of publicly available FFL‐MPI datasets has hindered algorithm development and system performance optimization. To address this issue, this paper presents a dataset based on the FFL‐MPI system, named FFL‐MPI‐Data, through simulation studies and data acquisition from a real‐world system. Acquisition and Validation Methods The FFL‐MPI‐Data contains both simulated and measured data spanning multiple domains and harmonics. The measured dataset was acquired using a self‐developed scanner, which collected raw voltage signals, sinograms, and reconstructed images from 100 physical phantoms. The simulated dataset was generated via a custom‐built simulation framework that replicates the key parameters of the experimental scanner, with the MNIST dataset employed as digital phantoms. The simulation framework is validated by evaluating the characteristics of harmonic images output from numerical phantoms. Furthermore, the utility and reliability of the dataset were verified by evaluating the performance of a deep learning model (FBPConvNet) on both sparse‐angle and limited‐angle reconstruction tasks. Data Format and Usage Notes The FFL‐MPI‐Data repository adopts the Hierarchical Data Format (HDF5) for data storage, which contains both phantom data and associated metadata. The dataset is publicly available on the Zenodo platform. Furthermore, we provide open‐source codes for data processing, image reconstruction, simulation implementation, and deep learning demonstrations at https://github.com/BUAALGH/FFL‐MPI‐Data . Users can readily access and utilize the FFL‐MPI‐Data dataset via the provided example code. Potential Applications FFL‐MPI‐Data represents a publicly accessible dataset for FFL‐MPI, addressing the scarcity of available data for projection imaging research in this domain. This dataset is expected to facilitate the development and validation of image reconstruction algorithms, including investigations into signal processing, multi‐domain reconstruction, and multi‐harmonic reconstruction for specific tasks such as sparse‐angle and limited‐angle reconstruction.
Abstract Background Spatiotemporal fractionation redistributes dose between fractions and may reduce normal‐tissue exposure. Its sensitivity to setup uncertainty and radiobiological assumptions requires evaluation before clinical use. Purpose To evaluate proliferation‐based constrained spatiotemporal fractionation (p‐cSTF) under geometric and radiobiological perturbations in conventionally fractionated liver radiotherapy. The focus was normal‐liver dose. Methods Thirty‐five patients with hepatocellular carcinoma were included. Uniform fractionation, proliferation‐based spatiotemporal fractionation (p‐STF), and p‐cSTF were evaluated over 25 fractions. The primary comparison matched nominal mean gross tumor volume (GTV) biologically effective dose (BED); a supplementary comparison matched nominal model‐derived tumor control probability (TCP). Setup tests comprised 18 single‐axis offsets (± 1, ± 2, and ± 3 mm) and 24 transverse voxel offsets, each held constant over treatment. Thirty scenarios varied biological‐model parameters; precision checks used two 480‐scenario sequences. Nominal dose grids were evaluated under perturbations without reoptimization. Endpoints included physical dose, BED, equivalent dose in 2 Gy fractions (EQD2), coverage, and response‐model scores. Results Under biological‐parameter variation, normal‐liver (liver excluding GTV) BED was 26.22 ± 7.00 Gy for p‐cSTF, 28.18 ± 7.57 Gy for Uniform, and 28.05 ± 7.76 Gy for p‐STF. Corresponding EQD2 values were 14.92 ± 3.96, 16.05 ± 4.29, and 15.96 ± 4.39 Gy. Values are cohort means and between‐patient standard deviations of patient‐specific scenario means. Normal‐liver BED remained lower with p‐cSTF after nominal TCP matching, although fixed transverse offsets produced lower GTV TCP than with Uniform. Expanded sampling retained the direction of the liver‐dose difference; the original 30 scenarios did not meet the specified precision criteria for scenario variability. Conclusions P‐cSTF retained a simulated normal‐liver dose advantage under the tested perturbations, including after nominal TCP matching. This advantage must be considered alongside target‐response trade‐offs and uncertainty in scenario‐variability estimates. Clinical toxicity and treatment deliverability were not evaluated.
