Journal of Medical Imaging最新文献

Purpose: Our study investigates the potential benefits of incorporating prior anatomical knowledge into a deep learning (DL) method designed for the automated segmentation of lung lobes in chest CT scans.

Approach: We introduce an automated DL-based approach that leverages anatomical information from the lung's vascular system to guide and enhance the segmentation process. This involves utilizing a lung vessel connectivity (LVC) map, which encodes relevant lung vessel anatomical data. Our study explores the performance of three different neural network architectures within the nnU-Net framework: a standalone U-Net, a multitasking U-Net, and a cascade U-Net.

Results: Experimental findings suggest that the inclusion of LVC information in the DL model can lead to improved segmentation accuracy, particularly, in the challenging boundary regions of expiration chest CT volumes. Furthermore, our study demonstrates the potential for LVC to enhance the model's generalization capabilities. Finally, the method's robustness is evaluated through the segmentation of lung lobes in 10 cases of COVID-19, demonstrating its applicability in the presence of pulmonary diseases.

Conclusions: Incorporating prior anatomical information, such as LVC, into the DL model shows promise for enhancing segmentation performance, particularly in the boundary regions. However, the extent of this improvement has limitations, prompting further exploration of its practical applicability.

Purpose: Segmentation of ovarian/adnexal masses from surrounding tissue on ultrasound images is a challenging task. The separation of masses into different components may also be important for radiomic feature extraction. Our study aimed to develop an artificial intelligence-based automatic segmentation method for transvaginal ultrasound images that (1) outlines the exterior boundary of adnexal masses and (2) separates internal components.

Approach: A retrospective ultrasound imaging database of adnexal masses was reviewed for exclusion criteria at the patient, mass, and image levels, with one image per mass. The resulting 54 adnexal masses (36 benign/18 malignant) from 53 patients were separated by patient into training (26 benign/12 malignant) and independent test (10 benign/6 malignant) sets. U-net segmentation performance on test images compared to expert detailed outlines was measured using the Dice similarity coefficient (DSC) and the ratio of the Hausdorff distance to the effective diameter of the outline ( $R_{\mathrm{HD}\text{-}D}$ ) for each mass. Subsequently, in discovery mode, a two-level fuzzy c-means (FCM) unsupervised clustering approach was used to separate the pixels within masses belonging to hypoechoic or hyperechoic components.

Results: The DSC (median [95% confidence interval]) was 0.91 [0.78, 0.96], and $R_{\mathrm{HD}\text{-}D}$ was 0.04 [0.01, 0.12], indicating strong agreement with expert outlines. Clinical review of the internal separation of masses into echogenic components demonstrated a strong association with mass characteristics.

Conclusion: A combined U-net and FCM algorithm for automatic segmentation of adnexal masses and their internal components achieved excellent results compared with expert outlines and review, supporting future radiomic feature-based classification of the masses by components.

Purpose: Endometrial cancer (EC) is one of the most common types of cancer affecting women. While the hematoxylin-and-eosin (H&E) staining remains the standard for histological analysis, the immunohistochemistry (IHC) method provides molecular-level visualizations. Our study proposes a digital staining method to generate the hematoxylin-3,3'-diaminobenzidine (H-DAB) IHC stain of Ki-67 for the whole slide image of the EC tumor from its H&E stain counterpart.

Approach: We employed a color unmixing technique to yield stain density maps from the optical density (OD) of the stains and utilized the U-Net for end-to-end inference. The effectiveness of the proposed method was evaluated using the Pearson correlation between the digital and physical stain's labeling index (LI), a key metric indicating tumor proliferation. Two different cross-validation schemes were designed in our study: intraslide validation and cross-case validation (CCV). In the widely used intraslide scheme, the training and validation sets might include different regions from the same slide. The rigorous CCV validation scheme strictly prohibited any validation slide from contributing to training.

Results: The proposed method yielded a high-resolution digital stain with preserved histological features, indicating a reliable correlation with the physical stain in terms of the Ki-67 LI. In the intraslide scheme, using intraslide patches resulted in a biased accuracy (e.g., $R=0.98$ ) significantly higher than that of CCV. The CCV scheme retained a fair correlation (e.g., $R=0.66$ ) between the LIs calculated from the digital stain and its physical IHC counterpart. Inferring the OD of the IHC stain from that of the H&E stain enhanced the correlation metric, outperforming that of the baseline model using the RGB space.

Conclusions: Our study revealed that molecule-level insights could be obtained from H&E images using deep learning. Furthermore, the improvement brought via OD inference indicated a possible method for creating more generalizable models for digital staining via per-stain analysis.

Purpose: To help radiologists examine the growing number of computed tomography (CT) scans, automatic anomaly detection is an ongoing focus of medical imaging research. Radiologists must analyze a CT scan by searching for any deviation from normal healthy anatomy. We propose an approach to detecting abnormalities in axial 2D CT slice images of the brain. Although much research has been done on detecting abnormalities in magnetic resonance images of the brain, there is little work on CT scans, where abnormalities are more difficult to detect due to the low image contrast that must be represented by the model used.

Approach: We use a generative adversarial network (GAN) to learn normal brain anatomy in the first step and compare two approaches to image reconstruction: training an encoder in the second step and using iterative optimization during inference. Then, we analyze the differences from the original scan to detect and localize anomalies in the brain.

Results: Our approach can reconstruct healthy anatomy with good image contrast for brain CT scans. We obtain median Dice scores of 0.71 on our hemorrhage test data and 0.43 on our test set with additional tumor images from publicly available data sources. We also compare our models to a state-of-the-art autoencoder and a diffusion model and obtain qualitatively more accurate reconstructions.

Conclusions: Without defining anomalies during training, a GAN-based network was used to learn healthy anatomy for brain CT scans. Notably, our approach is not limited to the localization of hemorrhages and tumors and could thus be used to detect structural anatomical changes and other lesions.

Purpose: The measures that traditionally describe the levator hiatus (LH) are straightforward and reliable; however, they were not specifically designed to capture significant differences. Statistical shape modeling (SSM) was used to quantify LH shape variation across reproductive-age women and identify novel variables associated with LH size and shape.

Approach: A retrospective study of pelvic MRIs from 19 nulliparous, 32 parous, and 12 pregnant women was performed. The LH was segmented in the plane of minimal LH dimensions. SSM was implemented. LH size was defined by the cross-sectional area, maximal transverse diameter, and anterior-posterior (A-P) diameter. Novel SSM-guided variables were defined by regions of greatest variation. Multivariate analysis of variance (MANOVA) evaluated group differences, and correlations determined relationships between size and shape variables.

Results: Overall shape ( $p<0.001$ ), SSM mode 2 (oval to $T$ -shape, $p=0.002$ ), mode 3 (rounder to broader anterior shape, $p=0.004$ ), and maximal transverse diameter ( $p=0.003$ ) significantly differed between groups. Novel anterior and posterior transverse diameters were identified at 14% and 79% of the A-P length. Anterior transverse diameter and maximal transverse diameter were strongly correlated ( $r=0.780$ , $p<0.001$ ), while posterior transverse diameter and maximal transverse diameter were weakly correlated ( $r=0.398$ , $p=0.001$ ).

Conclusions: The traditional maximal transverse diameter generally corresponded with SSM findings but cannot describe anterior and posterior variation independently. The novel anterior and posterior transverse diameters represent both size and shape variation, can be easily calculated alongside traditional measures, and are more sensitive to subtle and local LH variation. Thus, they have a greater ability to serve as predictive and diagnostic parameters.

Purpose: Radiologists are tasked with visually scrutinizing large amounts of data produced by 3D volumetric imaging modalities. Small signals can go unnoticed during the 3D search because they are hard to detect in the visual periphery. Recent advances in machine learning and computer vision have led to effective computer-aided detection (CADe) support systems with the potential to mitigate perceptual errors.

Approach: Sixteen nonexpert observers searched through digital breast tomosynthesis (DBT) phantoms and single cross-sectional slices of the DBT phantoms. The 3D/2D searches occurred with and without a convolutional neural network (CNN)-based CADe support system. The model provided observers with bounding boxes superimposed on the image stimuli while they looked for a small microcalcification signal and a large mass signal. Eye gaze positions were recorded and correlated with changes in the area under the ROC curve (AUC).

Results: The CNN-CADe improved the 3D search for the small microcalcification signal ( $\Delta \mathrm{AUC}=0.098$ , $p=0.0002$ ) and the 2D search for the large mass signal ( $\Delta \mathrm{AUC}=0.076$ , $p=0.002$ ). The CNN-CADe benefit in 3D for the small signal was markedly greater than in 2D ( $\Delta \Delta \mathrm{AUC}=0.066$ , $p=0.035$ ). Analysis of individual differences suggests that those who explored the least with eye movements benefited the most from the CNN-CADe ( $r=-0.528$ , $p=0.036$ ). However, for the large signal, the 2D benefit was not significantly greater than the 3D benefit ( $\Delta \Delta \mathrm{AUC}=0.033$ , $p=0.133$ ).

Conclusion: The CNN-CADe brings unique performance benefits to the 3D (versus 2D) search of small signals by reducing errors caused by the underexploration of the volumetric data.

Purpose: Analyzing the anatomy of the aorta and left ventricular outflow tract (LVOT) is crucial for risk assessment and planning of transcatheter aortic valve implantation (TAVI). A comprehensive analysis of the aortic root and LVOT requires the extraction of the patient-individual anatomy via segmentation. Deep learning has shown good performance on various segmentation tasks. If this is formulated as a supervised problem, large amounts of annotated data are required for training. Therefore, minimizing the annotation complexity is desirable.

Approach: We propose two-dimensional (2D) cross-sectional annotation and point cloud-based surface reconstruction to train a fully automatic 3D segmentation network for the aortic root and the LVOT. Our sparse annotation scheme enables easy and fast training data generation for tubular structures such as the aortic root. From the segmentation results, we derive clinically relevant parameters for TAVI planning.

Results: The proposed 2D cross-sectional annotation results in high inter-observer agreement [Dice similarity coefficient (DSC): 0.94]. The segmentation model achieves a DSC of 0.90 and an average surface distance of 0.96 mm. Our approach achieves an aortic annulus maximum diameter difference between prediction and annotation of 0.45 mm (inter-observer variance: 0.25 mm).

Conclusions: The presented approach facilitates reproducible annotations. The annotations allow for training accurate segmentation models of the aortic root and LVOT. The segmentation results facilitate reproducible and quantifiable measurements for TAVI planning.

Purpose: As large analyses merge data across sites, a deeper understanding of variance in statistical assessment across the sources of data becomes critical for valid analyses. Diffusion tensor imaging (DTI) exhibits spatially varying and correlated noise, so care must be taken with distributional assumptions. Here, we characterize the role of physiology, subject compliance, and the interaction of the subject with the scanner in the understanding of DTI variability, as modeled in the spatial variance of derived metrics in homogeneous regions.

Approach: We analyze DTI data from 1035 subjects in the Baltimore Longitudinal Study of Aging, with ages ranging from 22.4 to 103 years old. For each subject, up to 12 longitudinal sessions were conducted. We assess the variance of DTI scalars within regions of interest (ROIs) defined by four segmentation methods and investigate the relationships between the variance and covariates, including baseline age, time from the baseline (referred to as "interval"), motion, sex, and whether it is the first scan or the second scan in the session.

Results: Covariate effects are heterogeneous and bilaterally symmetric across ROIs. Inter-session interval is positively related ( $p\ll 0.001$ ) to FA variance in the cuneus and occipital gyrus, but negatively ( $p\ll 0.001$ ) in the caudate nucleus. Males show significantly ( $p\ll 0.001$ ) higher FA variance in the right putamen, thalamus, body of the corpus callosum, and cingulate gyrus. In 62 out of 176 ROIs defined by the Eve type-1 atlas, an increase in motion is associated ( $p<0.05$ ) with a decrease in FA variance. Head motion increases during the rescan of DTI ( $\Delta \mu =0.045$ mm per volume).

Conclusions: The effects of each covariate on DTI variance and their relationships across ROIs are complex. Ultimately, we encourage researchers to include estimates of variance when sharing data and consider models of heteroscedasticity in analysis. This work provides a foundation for study planning to account for regional variations in metric variance.

Purpose: Optical coherence tomography (OCT) is an emerging imaging tool in healthcare with common applications in ophthalmology for detection of retinal diseases, as well as other medical domains. The noise in OCT images presents a great challenge as it hinders the clinician's ability to diagnosis in extensive detail.

Approach: In this work, a region-based, deep-learning, denoising framework is proposed for adaptive cleaning of noisy OCT-acquired images. The core of the framework is a hybrid deep-learning model named transformer enhanced autoencoder rendering (TEAR). Attention gates are utilized to ensure focus on denoising the foreground and to remove the background. TEAR is designed to remove the different types of noise artifacts commonly present in OCT images and to enhance the visual quality.

Results: Extensive quantitative evaluations are performed to evaluate the performance of TEAR and compare it against both deep-learning and traditional state-of-the-art denoising algorithms. The proposed method improved the peak signal-to-noise ratio to 27.9 dB, CNR to 6.3 dB, SSIM to 0.9, and equivalent number of looks to 120.8 dB for a dental dataset. For a retinal dataset, the performance metrics in the same sequence are: 24.6, 14.2, 0.64, and 1038.7 dB, respectively.

Conclusions: The results show that the approach verifiably removes speckle noise and achieves superior quality over several well-known denoisers.