Journal of Medical Imaging最新文献

Purpose: Connectome network metrics are commonly regarded as fundamental properties of the brain, and their alterations have been implicated in the development of Alzheimer's disease, multiple sclerosis, and traumatic brain injury. However, these metrics are actually estimated properties through a multistage propagation from local voxel diffusion estimations, regional tractography, and region of interest mapping. These estimation processes are significantly influenced by choices specific to imaging protocols and software, producing site-wise effects.

Approach: Recent advances in disentanglement techniques offer opportunities to learn representational spaces that separate factors that cause domain shifts from intrinsic biological factors. Although these techniques have been applied in unsupervised brain anomaly detection and image-level features, their application to the unique manifold structures of connectome adjacency matrices remains unexplored. Here, we explore the conditional variational autoencoder structure for generating site-invariant representations of the connectome, allowing the harmonization of brain network measures.

Results: Focusing on the context of aging, we conducted a study involving 823 patients across two sites. This approach effectively segregates site-specific influences from biological features, aligns network measures across different domains (Cohen's $D<0.2$ and Mann-Whitney $U\text{-}\text{test}<0.05$ ), and maintains associations with age ( $2.71\times {10}^{-02}\pm 2.86\times {10}^{-03}$ error in years) and sex ( $0.92\pm 0.02$ accuracy).

Conclusions: Our findings demonstrate that using latent representations significantly harmonizes network measures and provides robust metrics for multi-site brain network analysis.

Purpose: The prevalence of type 2 diabetes mellitus (T2DM) has been steadily increasing over the years. We aim to predict the occurrence of T2DM using mammography images within 5 years using two different methods and compare their performance.

Approach: We examined 312 samples, including 110 positive cases (developed T2DM after 5 years) and 202 negative cases (did not develop T2DM) using two different methods. In the first method, a radiomics-based approach, we utilized radiomics features and machine learning (ML) algorithms. The entire breast region was chosen as the region of interest for extracting radiomics features. Then, a binary breast image was created from which we extracted 668 features and analyzed them using various ML algorithms. In the second method, a complex convolutional neural network (CNN) with a modified ResNet architecture and various kernel sizes was applied to raw mammography images for the prediction task. A nested, stratified five-fold cross-validation was done for both parts A and B to compute accuracy, sensitivity, specificity, and area under the receiver operating curve (AUROC). Hyperparameter tuning was also done to enhance the model's performance and reliability.

Results: The radiomics approach's light gradient boosting model gave 68.9% accuracy, 30.7% sensitivity, 89.5% specificity, and 0.63 AUROC. The CNN method achieved an AUROC of 0.58 over 20 epochs.

Conclusion: Radiomics outperformed CNN by 0.05 in terms of AUROC. This may be due to the more straightforward interpretability and clinical relevance of predefined radiomics features compared with the complex, abstract features learned by CNNs.

Purpose: Diffusion magnetic resonance imaging (dMRI) quantitatively estimates brain microstructure, diffusion tractography being one clinically utilized framework. To advance such dMRI approaches, direct quantitative comparisons between microscale anisotropy and orientation are imperative. Complete backscattering Mueller matrix polarized light imaging (PLI) enables the imaging of thin and thick tissue specimens to acquire numerous optical metrics not possible through conventional transmission PLI methods. By comparing complete PLI to dMRI within the ferret optic chiasm (OC), we may investigate the potential of this PLI technique as a dMRI validation tool and gain insight into the microstructural and orientational sensitivity of this imaging method in different tissue thicknesses.

Approach: Post-mortem ferret brain tissue samples (whole brain, $n=1$ and OC, $n=3$ ) were imaged with both dMRI and complete backscattering Mueller matrix PLI. The specimens were sectioned and then reimaged with PLI. Region of interest and correlation analyses were performed on scalar metrics and orientation vectors of both dMRI and PLI in the coherent optic nerve and crossing chiasm.

Results: Optical retardance and dMRI fractional anisotropy showed similar trends between metric values and were strongly correlated, indicating a bias to macroscale architecture in retardance. Thick tissue displays comparable orientation between the diattenuation angle and dMRI fiber orientation distribution glyphs that are not evident in the retardance angle.

Conclusions: We demonstrate that backscattering Mueller matrix PLI shows potential as a tool for microstructural dMRI validation in thick tissue specimens. Performing complete polarimetry can provide directional characterization and potentially microscale anisotropy information not available by conventional PLI alone.

Purpose: Digital phantoms are one of the key components of virtual imaging trials (VITs) that aim to assess and optimize new medical imaging systems and algorithms. However, these phantoms vary in their voxel resolution, appearance, and structural details. We investigate whether and how variations between digital phantoms influence system optimization with digital breast tomosynthesis (DBT) as a chosen modality.

Methods: We selected widely used and open-access digital breast phantoms created with different methods and generated an ensemble of DBT images to test acquisition strategies. Human observer performance was evaluated using localization receiver operating characteristic (LROC) studies for each phantom type. Noise power spectrum and gaze metrics were also employed to compare phantoms and generated images.

Results: Our LROC results show that the arc samplings for peak performance were $\sim 2.5\mathrm{deg}$ and 6 deg in Bakic and XCAT breast phantoms, respectively, for the 3-mm lesion detection task and indicate that system optimization outcomes from VITs can vary with phantom types and structural frequency components. In addition, a significant correlation ( $p<0.01$ ) between gaze metrics and diagnostic performance suggests that gaze analysis can be used to understand and evaluate task difficulty in VITs.

Conclusion: Our results point to the critical need to evaluate realism in digital phantoms and ensure sufficient structural variations at spatial frequencies relevant to the intended task. Standardizing phantom generation and validation tools may help reduce discrepancies among independently conducted VITs for system or algorithmic optimizations.

Purpose: We evaluate the impact of charge summing correction on a cadmium telluride (CdTe)-based photon-counting detector in breast computed tomography (CT).

Approach: We employ a custom-built laboratory benchtop system using the X-THOR FX30 0.75-mm CdTe detector (Varex Imaging, Salt Lake City, Utah, United States) with a pixel pitch of 0.1 mm, operated in both standard mode [single pixel (SP)] and charge summing correction mode [anticoincidence (AC)]. A tungsten anode source operated at 55 kVp with 2-mm aluminum external filtration and tube currents of 25, 100, and 200 mA with corresponding exposure times of 20, 5, and 2.5 ms were employed to study the effects of X-ray fluence and pulse pileup. Performance comparisons between AC and SP modes are performed in both projection and image reconstructed spaces. In the projection space, performance metrics include count rate, energy resolution, uniformity, modulation transfer function (MTF), and noise power spectrum (NPS). In the image space, performance metrics consist of contrast-to-noise ratio (CNR), uniformity, NPS, and iodine quantification accuracy. For both acquisition modes, signal-to-thickness calibration, for gain and beam hardening corrections, is used before image reconstruction. Images are reconstructed via TIGRE CT software using the standard Feldkamp, Davis, and Kress (FDK) filtered back projection algorithm with a Hann filter and reconstructed with a voxel size of 0.081 mm. Material decomposition is performed using a standard image-based method.

Results: In the detector space, the application of hardware-based charge summing correction enhances spectral resolution and improves the spatial resolution of MTF at lower energy thresholds but introduces anomalous edge enhancement effects and artifacts in the MTF at high fluence. A negative noise correlation was observed in AC mode-acquired images. As expected, the AC acquisition mode results in a decreased detector count rate. In the image space, NPS results displayed elevated noise in low-energy AC images. However, at high energy, noise was comparable between both modes. Greater uniformity was observed in SP mode-acquired images. The largest disparity was observed in the iodine quantification test, where the AC mode demonstrates a much stronger linear relationship between estimated and true iodine concentrations than the SP mode.

Conclusion: The results are specific to the studied system, reconstruction parameters, and irradiation conditions limited to 200 mA and 0.5 mAs. The AC mode generally provides better energy and MTF resolution at low energy thresholds but with increased noise and reduced uniformity. In image space, charge summing correction improved iodine quantification and CNR at high energy thresholds.

JMI Editor in Chief Bennett Landman provides an overview of JMI Volume 12 Issue 1 and spotlights key aspects of JMI peer review, with an eye toward continued growth for the JMI community.

Purpose: We aim to investigate the characteristics and evaluate the performance of synthetic mammograms (SMs) based on wide-angle digital breast tomosynthesis (DBT) compared with digital mammography (DM).

Approach: Fifty cases with both synthetic and digital mammograms were selected from the Malmö Breast Tomosynthesis Screening Trial. They were categorized into five groups consisting of normal cases and recalled cases with false-positive and true-positive findings from DM and DBT only. The DBT system used was a wide-angle (WA) system from Siemens, and the SM images were reconstructed from the DBT images. Visual grading, detection, and recall were evaluated by experienced breast radiologists in both SM and DM images.

Results: Some image quality criteria of the SM images were rated as qualitatively inferior to DM images. However, reader-averaged diagnostic accuracy (0.57 versus 0.55), sensitivity (0.46 versus 0.50), and specificity (0.64 versus 0.58) were not significantly different between SM and DM, respectively.

Conclusions: Synthetic mammography plays a promising role to complement or even replace DM. The study could not find any indications of substantial differences in the sensitivity or specificity of SM for WA DBT systems compared with DM. However, certain image quality criteria of SM fall slightly short compared with DM images. Next-generation DBT systems could address such limitations through improved reconstruction algorithms and system design, and their performance should be the focus of future research studies.

Purpose: Wireless capsule endoscopy (WCE) is a non-invasive technology used for diagnosing gastrointestinal abnormalities. A single examination generates $\sim \mathrm{55,000}$ images, making manual review both time-consuming and costly for doctors. Therefore, the development of computer vision-assisted systems is highly desirable to aid in the diagnostic process.

Approach: We presents a deep learning approach leveraging knowledge distillation (KD) from a convolutional neural network (CNN) teacher model to a vision transformer (ViT) student model for gastrointestinal abnormality recognition. The CNN teacher model utilizes attention mechanisms and depth-wise separable convolutions to extract features from WCE images, supervising the ViT in learning these representations.

Results: The proposed method achieves accuracy of 97% and 96% on the Kvasir and KID datasets, respectively, demonstrating its effectiveness in distinguishing normal from abnormal regions and bleeding from non-bleeding cases. The proposed approach offers computational efficiency and generalization to unseen datasets, outperforming several state-of-the-art methods.

Conclusions: We proposed a deep learning approach utilizing CNNs and a ViT with KD to effectively classify gastrointestinal diseases in WCE images. It demonstrates promising performance on public datasets, distinguishing normal from abnormal regions and bleeding from non-bleeding cases while offering optimal computational efficiency compared with existing methods, making it suitable for GI disease applications.

Purpose: Retrieving images of organs and their associated pathologies is essential for evidence-based clinical diagnosis. Deep neural hashing (DNH) has demonstrated the ability to retrieve images fast on large datasets. Conventional pairwise DNH methods can focus on semantic similarity between either organs or pathology of an image pair but not on both simultaneously.

Approach: We propose an organ and pathology contextual-supervised hashing approach (OPHash) learned using three types of samples (called bags) to learn accurate hash representation. Because only semantic similarity is inadequate to incorporate with these bags, we introduce relational similarity to generate identical hash codes from most similar image pairs. OPHash is trained by minimizing classification loss, two retrieval losses implemented using Cauchy cross-entropy and maximizing discriminator loss over training samples.

Results: Experiments are performed with two radiology datasets derived from the publicly available datasets. OPHash achieves 24% higher mean average precision than the state-of-the-art for top-100 retrieval.

Conclusion: OPHash retrieves images with semantic similarity of organs and their associated pathology. It is agnostic to image size as well. This method improves retrieval efficiency across diverse medical imaging datasets, accommodating multiple organs and pathologies. The code is available at https://github.com/asimmanna17/OPHash.

Purpose: We aim to investigate the localization, visibility, and measurement of lung nodules in digital chest tomosynthesis (DTS).

Approach: Computed tomography (CT), maximum intensity projections (CT-MIP) (transaxial versus coronal orientation), and computer-aided detection (CAD) were used as location reference, and inter- and intra-observer agreement regarding lung nodule size was assessed. Five radiologists analyzed DTS and CT images from 24 participants with lung $\text{nodules}\ge 100{\mathrm{mm}}^3$ , focusing on lung nodule localization, visibility, and measurement on DTS. Visual grading was used to compare if coronal or transaxial CT-MIP better facilitated the localization of lung nodules in DTS.

Results: The majority of the lung nodules (79%) were rated as visible in DTS, although less clearly in comparison with CT. Coronal CT-MIP was the preferred orientation in the task of locating nodules on DTS. On DTS, area-based lung nodule size estimates resulted in significantly less measurement variability when compared with nodule size estimated based on mean diameter (mD) ( $p<0.05$ ). Also, on DTS, area-based lung nodule size estimates were more accurate ( $\mathrm{SEE}=38.7{\mathrm{mm}}^3$ ) than lung nodule size estimates based on mean diameter ( $\mathrm{SEE}=42.7{\mathrm{mm}}^3$ ).

Conclusions: Coronal CT-MIP images are superior to transaxial CT-MIP images in facilitating lung nodule localization in DTS. Most $\text{nodules}\ge 100{\mathrm{mm}}^3$ found on CT can be visualized, correctly localized, and measured in DTS, and area-based measurement may be the key to more precise and less variable nodule measurements on DTS.