Journal of Medical Imaging最新文献

Purpose: We present a deep-learning-based methodology for estimating deformation in 2D echocardiograms. The goal is to automatically estimate the longitudinal strain of the left ventricle (LV) walls in images affected by speckle noise and acoustic occlusions.

Approach: The proposed methodology integrates algorithms for converting sparse to dense flow, a Res-UNet architecture for automatic myocardium segmentation, flow estimation using a global motion aggregation network, and the computation of longitudinal strain curves and the global longitudinal strain (GLS) index. The approach was evaluated using two echocardiographic datasets in apical four-chamber view, both modified with noise and acoustic shadows. The CAMUS dataset ( $N=250$ ) was used for LV wall segmentation, whereas a synthetic image database ( $N=2037$ ) was employed for flow estimation.

Results: Among the main performance metrics achieved are 98% [96 to 99] of correlation in the conversion from sparse to dense flow, a Dice index of $88.2\%\pm 3.8\%$ for myocardial segmentation, an endpoint error of 0.133 [0.13 to 0.14] pixels in flow estimation, and an error of 1.34% [0.94 to 2.09] in the estimation of the GLS index.

Conclusions: The results demonstrate improvements over previously reported performances while maintaining stability in echocardiograms with acoustic shadows. This methodology could be useful in clinical practice for the analysis of echocardiograms with noise artifacts and acoustic occlusions. Our code and trained models are publicly available at https://github.com/ArBioIIMAS/echo-gma.

Purpose: Personalized federated learning (PFL) has been explored to address data heterogeneity while preserving privacy, and its application in computer-aided detection/diagnosis (CAD) software has been investigated. Ditto, a commonly studied PFL method, trains global and personalized models but is limited by instability in model updates and high hyperparameter tuning costs. We proposed Improved Ditto, a PFL method that dynamically adjusts the proportion of global model weights during personalized model updates to enhance stability and reduce hyperparameter tuning costs.

Approach: We introduced a personalized model update rule in Improved Ditto that dynamically determines the proportion of global model weights based on the L2-norm of the gradient-derived and global-model-derived terms. This method was evaluated using three types of CAD software: cerebral aneurysm detection in magnetic resonance (MR) angiography images (segmentation), brain metastasis detection in contrast-enhanced T1-weighted MR images (object detection), and liver lesion classification in gadolinium-ethoxybenzyl-diethylenetriamine pentaacetic acid-enhanced MR images (classification). The proposed method was compared with several conventional methods.

Results: In two out of three CAD software, the performance of Improved Ditto was competitive with Ditto and other federated-learning-based methods. The proposed method achieved a narrower hyperparameter search space, which contributed to reducing the tuning costs. In addition, it improved the stability of personalized model updates, suggesting enhanced adaptability to diverse datasets and tasks.

Conclusions: We demonstrate that dynamically adjusting global model weights during personalized model updates can improve the stability and adaptability of PFL. The proposed method reduces the hyperparameter tuning costs and offers potential benefits for CAD software.

Purpose: Cochlear implants (CIs) are neural prosthetics used to treat patients with severe-to-profound hearing loss. Patient-specific modeling of CI stimulation of the auditory nerve fiber (ANF) can help audiologists improve the CI programming. These models require localization of the ANFs relative to the surrounding anatomy and the CI. Localization is challenging because the ANFs are so small that they are not directly visible in clinical imaging. We hypothesize that the position of the ANFs can be accurately inferred from the location of the internal auditory canal (IAC), which has high contrast in CT because the ANFs pass through this canal between the cochlea and the brain.

Approach: Inspired by VoxelMorph, we propose a deep atlas-based IAC segmentation network. We create a single atlas in which the IAC and ANFs are pre-localized. Our network is trained to produce deformation fields (DFs) mapping coordinates from the atlas to new target volumes and that accurately segment the IAC. We hypothesize that DFs that accurately segment the IAC in target images will also facilitate accurate atlas-based localization of the ANFs. As opposed to VoxelMorph, which aims to produce DFs that accurately register the entire volume, our contribution is an entirely self-supervised training scheme that aims to produce DFs that accurately segment the target structure. This self-supervision is facilitated using a loss function inspired by the Mumford-Shah functional. We call our method Deep Atlas-Based Segmentation using Mumford-Shah (DABS-MS).

Results: Results show that DABS-MS outperforms VoxelMorph for IAC segmentation. Tests with publicly available datasets for trachea and kidney segmentation also show significant improvement in segmentation accuracy, demonstrating the generalizability of the method.

Conclusions: Our proposed DABS-MS method can accurately segment the IAC, which can then facilitate the localization of the ANFs. This patient-specific modeling of CI stimulation of the ANFs can help audiologists improve the CI programming, leading to better outcomes for patients with severe-to-profound hearing loss.

The editorial introduces the Special Section on Medical Image Perception and Observer Performance for JMI Volume 12 Issue 5.

Purpose: Four-dimensional (4D) ultrasound imaging is widely used in clinics for diagnostics and therapy guidance. Accurate target tracking in 4D ultrasound is crucial for autonomous therapy guidance systems, such as radiotherapy, where precise tumor localization ensures effective treatment. Supervised deep learning approaches rely on reliable ground truth, making accurate labels essential. We investigate the reliability of expert-labeled ground truth data by evaluating intra- and inter-observer variability in landmark labeling for 4D ultrasound imaging in the liver.

Approach: Eight 4D liver ultrasound sequences were labeled by eight expert observers, each labeling eight landmarks three times. Intra- and inter-observer variability was quantified, and observer survey and motion analysis were conducted to determine factors influencing labeling accuracy, such as ultrasound artifacts and motion amplitude.

Results: The mean intra-observer variability ranged from $1.58\mathrm{mm}\pm 0.90\mathrm{mm}$ to $2.05\mathrm{mm}\pm 1.22\mathrm{mm}$ depending on the observer. The inter-observer variability for the two observer groups was $2.68\mathrm{mm}\pm 1.69\mathrm{mm}$ and $3.06\mathrm{mm}\pm 1.74\mathrm{mm}$ . The observer survey and motion analysis revealed that ultrasound artifacts significantly affected labeling accuracy due to limited landmark visibility, whereas motion amplitude had no measurable effect. Our measured mean landmark motion was $11.56\mathrm{mm}\pm 5.86\mathrm{mm}$ .

Conclusions: We highlight variability in expert-labeled ground truth data for 4D ultrasound imaging and identify ultrasound artifacts as a major source of labeling inaccuracies. These findings underscore the importance of addressing observer variability and artifact-related challenges to improve the reliability of ground truth data for evaluating target tracking algorithms in 4D ultrasound applications.

Purpose: The combination of multi-layer flat panel detector (FPDT) X-ray imaging and physics-based material decomposition algorithms allows for the removal of anatomical structures. However, the reliability of these algorithms may be compromised by unaccounted materials or scattered radiation.

Approach: We investigated the two-material decomposition performance of a multi-layer FPDT in the context of 2D chest radiography without and with a 13:1 anti-scatter grid employed. A matrix-based material decomposition (MBMD) (equivalent to weighted logarithmic subtraction), a matrix-based material decomposition with polynomial beam hardening pre-correction (MBMD-PBC), and a projection domain decomposition were evaluated. The decomposition accuracy of simulated data was evaluated by comparing the bone and soft tissue images to the ground truth using the structural similarity index measure (SSIM). Simulation results were supported by experiments using a commercially available triple-layer FPDT retrofitted to a digital X-ray system.

Results: Independent of the selected decomposition algorithm, uncorrected scatter leads to negative bone estimates, resulting in small SSIM values and bone structures to remain visible in soft tissue images. Even with a 13:1 anti-scatter grid employed, bone images continue to show negative bone estimates, and bone structures appear in soft tissue images. Adipose tissue on the contrary has an almost negligible effect.

Conclusions: In a contact scan, scattered radiation leads to negative bone contrast estimates in the bone images and remaining bone contrast in the soft tissue images. Therefore, accurate scatter estimation and correction algorithms are essential when aiming for material decomposition using image data obtained with a multi-layer FPDT.

Purpose: Artificial intelligence (AI)-based medical imaging devices often include lesion or organ segmentation capabilities. Existing methods for segmentation performance evaluation compare AI results with an aggregated reference standard using accuracy metrics such as the Dice coefficient or Hausdorff distance. However, these approaches are limited by lacking a gold standard and challenges in defining meaningful success criteria. To address this, we developed a statistical method to assess agreement between an AI device and multiple human experts without requiring a reference standard.

Approach: We propose a paired-testing method to evaluate whether an AI device's segmentation performance significantly differs from that of multiple human experts. The method compares device-to-expert dissimilarity with expert-to-expert dissimilarity, avoiding the need for a reference standard. We validated the method through (1) statistical simulations where the Dice coefficient performance is either shared ("overlap agreeable") or not shared ("overlap disagreeable") between the device and experts; (2) image-based simulations using 2D contours with shared or nonshared transformation parameters (transformation agreeable or disagreeable). We also applied the method to compare an AI segmentation algorithm with four radiologists using data from the Lung Image Database Consortium.

Results: Statistical simulations show the method controls type I error ( $\sim 0.05$ ) for overlap-agreeable and type II error ( $\sim 0$ ) for overlap-disagreeable scenarios. Image-based simulations show acceptable performance with a mean type I error of 0.07 (SD 0.03) for transformation-agreeable and a mean type II error of 0.07 (SD 0.18) for transformation-disagreeable cases.

Conclusions: The paired-testing method offers a new tool for assessing the agreement between an AI segmentation device and multiple human expert panelists without requiring a reference standard.

Purpose: Many deep generative models have been proposed to reconstruct pseudo-healthy images for anomaly detection. Among these models, the variational autoencoder (VAE) has emerged as both simple and efficient. Although significant progress has been made in refining the VAE within the field of computer vision, these advancements have not been extensively applied to medical imaging applications.

Approach: We present a benchmark that assesses the ability of multiple VAEs to reconstruct pseudo-healthy neuroimages for anomaly detection in the context of dementia. We first propose a rigorous methodology to define the optimal architecture of the vanilla VAE and select through random searches the best hyperparameters of the VAE variants. Relying on a simulation-based evaluation framework, we thoroughly assess the ability of 20 VAE models to reconstruct pseudo-healthy images for the detection of dementia-related anomalies in 3D brain ${}^{18}F$ -fluorodeoxyglucose (FDG) positron emission tomography (PET) and compare their performance.

Results: This benchmark demonstrated that the majority of the VAE models tested were able to reconstruct images of good quality and generate healthy-looking images from simulated images presenting anomalies.

Conclusions: Even if no model clearly outperformed all the others, the benchmark allowed identifying a few models that perform slightly better than the vanilla VAE. It further showed that many VAE-based models can generalize to the detection of anomalies of various intensities, shapes, and locations in 3D brain FDG PET.

Purpose: State space models have shown promise in medical image segmentation by modeling long-range dependencies with linear complexity. However, they are limited in their ability to capture local features, which hinders their capacity to extract multiscale details and integrate global and local contextual information effectively. To address these shortcomings, we propose the dual-path multi-scale Mamba UNet (DMM-UNet) model.

Approach: This architecture facilitates deep fusion of local and global features through multi-scale modules within a U-shaped encoder-decoder framework. First, we introduce the multi-scale channel attention selective scanning block in the encoder, which combines global selective scanning with multi-scale channel attention to model both long-range and local dependencies simultaneously. Second, we design the spatial attention selective scanning block for the decoder. This block integrates global scanning with spatial attention mechanisms, enabling precise aggregation of semantic features through gated weighting. Finally, we develop the multi-dimensional collaborative attention layer to extract complementary attention weights across height, width, and channel dimensions, facilitating cross-space-channel feature interactions.

Results: Experiments were conducted on the ISIC17, ISIC18, Synapse, and ACDC datasets. One of the indicators, Dice similarity coefficient, achieved 89.88% on the ISIC17 dataset, 90.52% on the ISIC18 dataset, 83.07% on the Synapse dataset, and 92.60% on the ACDC dataset. There are also other indicators that perform well on this model.

Conclusions: The DMM-UNet model effectively addresses the shortcomings of state space models by enabling the integration of both local and global features, improving segmentation performance, and offering enhanced multiscale feature fusion for medical image segmentation tasks.

Purpose: Perceptual error is a significant cause of medical errors in radiology. Given the amount of information in a medical image, an image interpreter may become distracted by information unrelated to their search pattern. This may be especially challenging for novices. We aim to examine teaching medical trainees to evaluate chest radiographs (CXRs) for pulmonary nodules on limited field-of-view (LFOV) images, with the field of view (FOV) restricted to the lungs and mediastinum.

Approach: Healthcare trainees with limited exposure to interpreting images were asked to identify pulmonary nodules on CXRs, half of which contained nodules. The control and experimental groups evaluated two sets of CXRs. After the first set, the experimental group was trained to evaluate LFOV images, and both groups were again asked to assess CXRs for pulmonary nodules. Participants were given surveys after this educational session to determine their thoughts about the training and symptoms of computer vision syndrome (CVS).

Results: There was a significant improvement in performance in pulmonary nodule identification for both the experimental and control groups, but the improvement was more considerable in the experimental group ( $p\text{-}\text{value}=0.022$ ). Survey responses were uniformly positive, and each question was statistically significant (all $p\text{-}\text{values}<0.001$ ).

Conclusions: Our results show that using LFOV images may be helpful when teaching trainees specific high-yield perceptual tasks, such as nodule identification. The use of LFOV images was associated with reduced symptoms of CVS.