Low-quality Images Research Articles

MPRAGElike: A novel approach to generate T1w images from multi-contrast gradient echo images for brain segmentation.

Brain segmentation and multi-parameter mapping (MPM) are important steps in neurodegenerative disease characterization. However, acquiring both a high-resolution T1w sequence like MPRAGE (standard input to brain segmentation) and an MPM in the same neuroimaging protocol increases scan time and patient discomfort, making it difficult to combine both in clinical examinations. A novel approach to synthesize T1w images from MPM images, named MPRAGElike, is proposed and compared to the standard technique used to produce synthetic MPRAGE images (synMPRAGE). Twenty-three healthy subjects were scanned with the same imaging protocol at three different 7T sites using universal parallel transmit RF pulses. SNR, CNR, and automatic brain segmentation results from both MPRAGElike and synMPRAGE were compared against an acquired MPRAGE. The proposed MPRAGElike technique produced higher SNR values than synMPRAGE for all regions evaluated while also having higher CNR values for subcortical structures. MPRAGE was still the image with the highest SNR values overall. For automatic brain segmentation, MPRAGElike outperformed synMPRAGE when compared to MPRAGE (median Dice Similarity Coefficient of 0.90 versus 0.29 and Average Asymmetric Surface Distance of 0.33 versus 2.93mm, respectively), in addition to being simple, flexible, and considerably more robust to low image quality than synMPRAGE. The MPRAGElike technique can provide a better and more reliable alternative to synMPRAGE as a substitute for MPRAGE, especially when automatic brain segmentation is of interest and scan time is limited.

A Novel Face Frontalization Method by Seamlessly Integrating Landmark Detection and Decision Forest into Generative Adversarial Network (GAN)

In real-world scenarios, posture variation and low-quality image resolution are two well-known factors that compromise the accuracy and reliability of face recognition system. These challenges can be overcome using various methods, including Generative Adversarial Networks (GANs). Despite this, concerns over the accuracy and reliability of GAN methods are increasing as the facial recognition market expands rapidly. The existing framework such as Two-Pathway GAN (TP-GAN) method has demonstrated that it is superior to numerous GAN methods that provide better face-texture details due to its unique deep neural network structure that allows it to perceive local details and global structure in a supervised manner. TP-GAN overcomes some of the obstacle associated with face frontalization tasks through the use of landmark detection and synthesis functions, but it remains challenging to achieve the desired performance across a wide range of datasets. To address the inherent limitations of TP-GAN, we propose a novel face frontalization method (NFF) combining landmark detection, decision forests, and data augmentation. NFF provides 2D landmark detection to integrate global structure with local details of the generator model so that more accurate facial feature representations and robust feature extractions can be achieved. NFF enhances the stability of the discriminator model over time by integrating decision forest capabilities into the TP-GAN discriminator core architecture that allows us to perform a wide range of facial pose tasks. Moreover, NFF uses data augmentation techniques to maximize training data by generating completely new synthetic data from existing data. Our evaluations are based on the Multi-PIE, FEI, and CAS-PEAL datasets. NFF results indicate that TP-GAN performance can be significantly enhanced by resolving the challenges described above, leading to high quality visualizations and rank-1 face identification.

A nonlinear sensitivity matrix for improving the imaging quality in electrical resistance tomography

Electrical Resistance Tomography (ERT) can image the conductivity distribution in a detection field by boundary measurements. The ERT imaging refers to a typical nonlinear inverse problem, and currently it is represented by a linear sensitivity matrix. However, the existing linearized sensitivity matrix cannot effectively represent the nonlinear effect in the imaging process, and leads to low-quality ERT image. This paper combines an integral equation on ERT with the Geselowitz-Lehr formula, and applies a Gaussian kernel function to derive a nonlinear sensitivity matrix. The new matrix can effectively reflect both linear and nonlinear characteristics in ERT, and has a clear theoretical basis and rationality. Essentially, the limitation of the existing linear sensitivity matrix is greatly addressed. Compared to four typical ERT imaging algorithms with the linear sensitivity matrix, the use of nonlinear sensitivity matrix can averagely decrease the relative error to 0.16, while can increase the correlation coefficient to 0.88 along with various noise levels. Hence, the proposed nonlinear sensitivity matrix can effectively improve ERT imaging quality.

Correlation of pulmonary artery systolic pressures derived from echocardiography and hemodynamic catheterization in patients with tricuspid regurgitation: a large single-center study

Abstract Introduction Echocardiography is the recommended modality for non-invasive assessment of pulmonary pressures. However, in patients with severe tricuspid regurgitation (TR), the reliability of echocardiographically derived systolic pulmonary artery pressure (sPAP) is questioned. Limited data exists on the correlation between echocardiography and right heart catheterization (RHC) in selected group of patients with severe TR. Purpose To evaluate the correlation between echocardiographic and RHC measurements for sPAP in patients with varying TR severity. Method Patients who underwent echocardiography and RHC within three days at a large single referral center were retrospectively included. Exclusion criteria included incomplete echocardiographic and RHC data, ≥moderate pulmonary stenosis, low-quality images, and pregnancy. Echocardiographic sPAP was calculated using simplified Bernoulli equation. Echocardiographic right atrial pressure (RAP) was assessed by evaluating inferior vena cava (IVC) dimension and collapsibility. A difference of ≥10 mmHg between the two modalities for sPAP was considered clinically significant. Results Of 3696 patients included, mean age was 64±14 years, 1689 (46%) were females, 1659 (45%) had systemic hypertension, and 582 (16%) had atrial fibrillation. TR was <moderate in 2475 (67%) patients, moderate in 499 (14%), moderate-severe in 277 (7%), and severe in 445 (12%). Pulmonary hypertension (PH) was diagnosed in 3079 subjects (83%). A good correlation was found between sPAP derived by echocardiographic and RHC in the whole cohort (correlation coefficient (LCC) 0.79, CI 0.78-0.80), in severe TR subgroup (LCC 0.78, CI 0.75-0.82), and in severe TR with PH subgroup (LCC 0.76, CI 0.71-0.79). The mean echocardiographic misestimation for sPAP in severe TR was 3.97 mmHg (CI -5.24 to -2.71). The correlation between the two modalities for RAP was moderate in the whole cohort (LCC 0.59, CI 0.57-0.61) and fair in severe TR (LCC 0.35, CI 0.27-0.42). A clinically significant difference between the two modalities for sPAP was found in 1504 patients (41%), predominantly due to the right ventricle-right atrium gradient misestimation, rather than RAP misestimation, in patients with no PH (67% vs 33%), patients with mild-moderate PH (mean Pulmonary Artery Pressure (mPAP) 21-34 mmHg) (67% vs 33%) and severe PH (mPAP ≥35 mmHg) (76% vs 24%). Conclusion The correlation between echocardiographic and RHC measurements is good for sPAP, even in severe TR patients, while it is only fair to moderate for RAP. A clinically significant difference between echocardiographic and RHC sPAP is more commonly observed in patients with PH, with right ventricle-right atrium gradient misestimation being the most common cause.

Impact of diabetes on coronary plaque characteristics and major adverse cardiovascular events

Abstract Background Cardiovascular disease (CVD) is a leading cause of global morbidity and mortality, with diabetes significantly exacerbating its progression. Characterizing coronary plaque composition and its association with cardiovascular outcomes is crucial for understanding disease pathogenesis. Understanding the influence of diabetes on these parameters can enhance our ability to manage CVD outcomes more effectively. Purpose We aimed to investigate the impact of diabetes on coronary plaque characteristics and its association with Major Adverse Cardiovascular Events (MACE) in a single-centre cohort. Methods We assessed 190 patients who underwent coronary computed tomography angiography (CCTA) for evaluation of chest pain, excluding 15 patients due to low image quality, and the remaining 175 were further evaluated. Statistical analyses included multiple regression models to examine the relationship between plaque characteristics and diabetes status, ANOVA to compare characteristics between groups, and logistic regression to calculate the odds ratio (OR) for MACE differences between diabetic and non-diabetic patients. MACE was defined as acute myocardial infarction, stroke, cardiovascular mortality, and revascularization. Results The study included 175 participants: 47 patients diagnosed with diabetes and 128 patients without. The mean age of the patients was 58.7 years (±11 years). A total of 270 coronary plaques were analyzed: 167 plaques from patients without diabetes and 103 plaques from patients with diabetes. The patient cohort consisted of 47.4% men. Patients without diabetes had significantly lower values compared to those with diabetes across several measures: the median (Interquartile Range - IQR) Agatston score (150.5 (55-349) vs 473.0 (179-1269), F(1, 254) = 27.909, p < 0.001, B = -527.715), lesion length (mm) (5.1 (3.5-7.1) vs 6,0 (4.1-11.2), F(1, 254) = 11.613, p < 0.001, B = -3.789), plaque burden (0.23 (0.16-0.32) vs 0.28 (0.20-0.38), F(1, 254) = 14.429, p < 0.001, B = -0.060), plaque volume (mm³) (13.2 (6.5-24.6) vs 20.8 (9.4-47.8), F(1, 254) = 16.267, p < 0.001, B = -26.315), fibrous volume (mm³) (4.6 (2.3-9.4) vs 7.7 (3.2-16.6), F(1, 254) = 21.238, p < 0.001, B = -8.930), fibrous fatty volume (mm³) (0.41 (0.09-0.9) vs 0.99 (0.2-2.3), F(1, 254) = 5.236, p = 0.023, B = -0.603) and dense calcium volume (mm³) (6.2 (3.1-13.9) vs 9.3 (3.5-23.4), F (1, 254) = 12.644, p < 0.001, B = -16.288). There were 16 MACEs in total, with diabetic patients showing significantly higher odds of experiencing MACE than non-diabetic patients (OR = 3.077, p = 0.035). Conclusions These findings suggest that diabetes is associated with more extensive and potentially more severe coronary artery disease, impacting the development of various plaque characteristics. Future research could explore larger cohorts to validate these findings and explore the underlying mechanisms linking diabetes, plaque characteristics, and cardiovascular outcomes.

Interobserver regional variability of left ventricular endocardium and epicardium delineation in echocardiography imaging

Abstract Introduction Accurate delineation of the left ventricle (LV) endocardium and epicardium remains a significant challenge in echocardiography due to low image quality and tissue contrast, leading to a compromised reproducibility of LV border tracing [1]. Whereas the global variability in echocardiography segmentation is extensively discussed in the literature [2], there is still a substantial gap in understanding the distribution of border delineation error across the different regions of LV. Purpose This study aims to extensively analyse discrepancies in the echocardiography delineation of LV endocardial and epicardial borders at global and regional levels by considering a large cohort of annotation experts. Methods LV endocardial and epicardial borders were independently delineated by 35 sonographers from multiple institutions in apical 4-chamber (A4C) and 2-chamber (A2C) echocardiography sequences at end-diastole (ED) and end-systole (ES) in a cohort of 10 patients, totalling 80 annotations per observer. The ES and ED frames were preselected, and their ground truth was determined by a clinical committee. LV sections were obtained via standard Simpson’s rule [3], yielding 82 reference points along the epicardium and endocardium, grouped into 4 regions (septal, lateral, anterior, inferior) across 3 levels (basal, mid, apical) plus the apex, resulting in 13 segments. Global segmentation error was quantified using the Dice Similarity Coefficient (DSC) and Hausdorff 95% distance (HD95). Regional error was estimated as the minimum Euclidean distance from each of the 82 ground truth reference points to the segmentation border, averaged across LV segments and patients. Statistical significance was assessed via paired t-test. Results The average DSC was 0.87±0.03 and 0.92±0.02, and the average HD95 was 6.0±1.4 and 5.8±1.1 mm for endocardium and epicardium, respectively (Figure 1). Segmentation error was significantly lower at ED compared to ES (p=0.003) and in A4C compared to A2C (p=0.001). In line with the latter, the error in the septal region was significantly the lowest (Figure 2). A positive error trend base-to-apex is observed, with the error significantly the smallest at the basal level. Errors were more prominent in the apical lateral endocardial segment in A4C and in the apical anterior segment in A2C. These errors can be explained by the higher LV curvature in these segments [4]. Epicardium apex error could have been slightly underestimated due to some of the contours falling outside of the field of view in this region. Conclusion This work introduces a reproducible method to analyse global and regional variability in LV border tracing in echocardiography. The study identifies compromised segments that can be targeted via enhanced training and standardised protocols to effectively reduce interobserver variability, leading to consistent and accurate interpretations, ultimately benefiting patient care.

Prognostic value of left atrial strain index (LASI) in patients with heart failure with preserved ejection fraction

Abstract Background Left atrial (LA) strain has been proposed as an additional echocardiographic parameter to conventional diastolic dysfunction assessment for the estimation of elevated left ventricular end-diastolic pressure (LVEDP). Recently, left atrial strain index (LASI) has been introduced, which integrates by a machine-learned algorithm the entire strain curve (not limited to either the reservoir or contractile function) into a single parameter, and which was able to distinguish elevated versus normal LVEDP upon validation with invasive measurements. However, the prognostic value of LASI has not been widely established in the clinical setting. Purpose To investigate the prognostic value of LASI in patients with heart failure with preserved ejection fraction (HFpEF). Methods A total of 211 diagnosed HFpEF patients from a dedicated outpatient program were retrospectively analysed. Patients with cardiac devices or low image quality for LA analyses were excluded. Extensive echocardiographic analysis was performed, and LASI was assessed: positive values indicating elevated LVEDP and negative values identifying normal LVEDP (Figure 1). Study outcome was a composite of heart failure hospitalisation and all-cause mortality. Results From the total population, 117 (55%) patients had elevated LVEDP based on LASI and 94 (45%) normal LVEDP. Patients with elevated LVEDP were slightly older (77±6 vs. 74±8 years, p=0.003), had more often atrial fibrillation (61% vs. 10%, p<0.001), higher HFA-PEFF score (HFA-PEFF≥5 69% vs. 53%, p=0.017) and NT-proBNP (135 [62-194] vs. 44 [25-81] pmol/L, p<0.001), and more impaired diastology parameters [LA indexed volume 57±20 vs. 44±12 ml/m2, p<0.001; average E/e’ 13±5 vs. 11±3, p=0.005; tricuspid jet velocity 2.8±0.4 vs 2.6±0.4 cm/s, p=0.036). During a median follow-up of 61 months, a total of 95 events were recorded. The elevated LVEDP group had worse outcome compared to the normal LVEDP group classified based on LASI (p<0.001, Figure 1). When stratifying first according to HFA-PEFF score (cut-off ≥5) and then into LVEDP groups, patients with the elevated LVEDP still had worse outcomes, regardless of HFA-PEFF score (Figure 2). Uni- and multivariable survival analysis demonstrated that the LASI classification was independently associated with outcome, after correcting for age, sex, NYHA class, comorbidities, kidney function and HFA-PEFF score (HR 2.716 [1.610-4.583], p<0.001). Likelihood analysis showed an incremental value upon addition of LASI to the clinical factors (ꭓ2 44 to 56, p<0.001). Subgroup analysis on patients in sinus rhythm also demonstrated similar association of LASI to outcomes, adding significant value to other parameters. Conclusion LASI is significantly associated with outcome in HFpEF patients, and can be an adjunct to current methods of diastolic dysfunction assessment to improve risk stratification of these patients.

Echocardiography image quality assessment: human subjectivity and artificial intelligence prediction in multi-centre data

Abstract Background Echocardiography image quality is important for precise care, but its assessment is subjective. An image quality metric by artificial intelligence (AI) may permit the optimization of image acquisition and improve analysis. Purpose To investigate (i) the most relevant features of echocardiography images related to perceived quality, (ii) to quantify them with AI and (iii) test their variability among centres. Methods Fifteen thousand anonymized transthoracic echocardiography studies acquired at Vall d’Hebron Hospital (VH) during care were retrieved. The original clinical assessment of study quality was obtained. Via co-creation, three image quality metrics (each with 3 levels, [0, 1, 2]) were identified: (i) the definition of structures borders and the presence of (ii) all expected cardiac structures and (iii) of apical foreshortening. Then, cardiologists annotated image quality of 6831 VH and 1000 CAMUS videos, which provided its own quality metric [1]. Three AI models were trained with 5256 videos from the VH to reproduce the 3 quality metrics. Models were then internally and externally validated in 1424 and 1000 videos from VH and CAMUS, respectively. Results Data from 488 VH patients were available for this analysis (Table 1). Inter-observer agreement in VH data was limited (56, 48 and 46% for border definition, presence of structures and foreshortening, respectively). Borders definition (p<0.001) but not the presence of all expected structures (p=0.862) nor apical foreshortening (p=0.440) was related to clinical study quality. Trained with data from a single rater, AI models demonstrated reasonable performance in the predictions for borders definition (agreement 60%), presence of all expected structures (54%) and apical foreshortening (53%) in VH images, showing slightly better performance when tested with data annotated by the same rater (62%, 58%, 51%). Border definition decreased with BMI (p=0.042), which was reproduced by the model (p=0.001), while annotated (p=0.028) but not predicted (p=0.525) border definition was lower in cardiomyopathy, a potential bias of the model. Age, sex, aortic stenosis and irregular heart rhythm did not show trends with border definition. In CAMUS data, annotated image quality metrics did not compare well with original annotations (agreement of 21% and 35% for border definition and complete structure), which showed lower image quality compared to VH (60 vs 11% of CAMUS videos defined as low quality by VH vs CAMUS raters), highlighting the high subjectivity of image quality. Generalizability to CAMUS images was good for border definition (agreement of 57%) but very limited for structures completeness (19%). Conclusions Echocardiography image quality is subjective and mainly driven by the definition of the borders of cardiac structures. AI models can learn image quality, but generalization to images from other centres may be limited. Demographic and clinical information

Deep learning aided determination of the optimal number of detectors for photoacoustic tomography.

Photoacoustic tomography (PAT) is a non-destructive, non-ionizing, and rapidly expanding hybrid biomedical imaging technique, yet it faces challenges in obtaining clear images due to limited data from detectors or angles. As a result, the methodology suffers from significant streak artifacts and low-quality images. The integration of deep learning (DL), specifically convolutional neural networks (CNNs), has recently demonstrated powerful performance in various fields of PAT. This work introduces a post-processing-based CNN architecture named residual-dense UNet (RDUNet) to address the stride artifacts in reconstructed PA images. The framework adopts the benefits of residual and dense blocks to form high-resolution reconstructed images. The network is trained with two different types of datasets to learn the relationship between the reconstructed images and their corresponding ground truths (GTs). In the first protocol, RDUNet (identified as RDUNet I) underwent training on heterogeneous simulated images featuring three distinct phantom types. Subsequently, in the second protocol, RDUNet (referred to as RDUNet II) was trained on a heterogeneous composition of 81% simulated data and 19% experimental data. The motivation behind this is to allow the network to adapt to diverse experimental challenges. The RDUNet algorithm was validated by performing numerical and experimental studies involving single-disk, T-shape, and vasculature phantoms. The performance of this protocol was compared with the famous backprojection (BP) and the traditional UNet algorithms. This study shows that RDUNet can substantially reduce the number of detectors from 100 to 25 for simulated testing images and 30 for experimental scenarios.

¡No Pasaran! The Image of Spain and the Image of the Enemy in a Soviet Information Photo Album of 1937

The paths of Russia and Spain crossed several times in the 20th century. The Spanish Civil War of 1936–1939 has always been and remains a topic that attracts the attention of experts and influences the development of a multifaceted cultural dialogue between the two countries. The events of those years contributed greatly to shaping the image of Spain in the Soviet Union. In recent decades, a number of works have been published, exploring not only the image of Spain in the USSR in the 1930s, but also the image of the enemy, which emerged as the first direct military confrontation with fascist regimes took place. An important topic analyzed in these works is the tools of visual securitization, which involves the formation of the enemy image by visual means. This paper is based on the materials included in the photo album ¡No pasaran! They will not pass! published by the Central Committee of the Komsomol in 1937. This source traditionally falls into the category of photo journalism. It is a type of source that is not examined much in literature, yet it is interesting as it combines visual images and verbal characteristics that reinforce stereotypical perceptions and key mythologems relevant to that period. The material illustrates how the image of Spain took shape in the USSR during the Civil War and how the image of the enemy emerged before World War II. The album was published as part of a pro-Spain campaign in the Soviet Union, when the Republican faction was still believed to win soon. The verbal content of the album is stronger than the images in terms of impact, although it is usually the visual content that is more impactful. Unlike poster art with its potential for visual securitization, a photo album, which combines extensive textual material with low-quality images, reinforces stereotypical ideas by combining these two sources of information, with the obvious dominance of the former.

Risk of clinically significant prostate cancer undercategorized by multiparametric magnetic resonance imaging.

To investigative potential clinicopathological characteristics and imaging-related risk factors of clinically significant prostate cancer (csPCa) undercategorized in patients with negative or equivocal MRI. This retrospective study included 581 patients with pathologically confirmed csPCa (Gleason score ≥ 3 + 4), including 108 undercategorized csPCa and 473 detected csPCa. All patients underwent multiparametric MRI (mpMRI). The undercategorized csPCa was defined as a MRI result with PI-RADS ≤ 3. The clinicopathological characteristics and imaging-related factors were compared between the undercategorized group(Group A) (PI-RADS 1-3) and detected group (Group B) (PI-RADS 4-5). The age, total PSA levels, PSAD, free PSA, prostate imaging quality (PI-QUAL) scores, and Gleason scores were significantly lower in the Group A than Group B. The lesions were larger and involved in peripheral and transition zones in the Group B. A significant difference in the second reading opinion. Age (odds ratio [OR], 0.94), PSAD (OR, 0.09), and PI-QUAL scores (OR, 0.25) were significantly associated with the undercategorized csPCa. The rate of undercategorized csPCa with these three risk factors (age, PSAD, and PI-QUAL scores of < 71, < 0.355, and < 3, respectively) was 68.62%. The lack of zoomed-DWI resulted in lower PI-QUAL scores. Finally, the probability of undercategorized csPCa without zoomed DWI was 3.186 times higher than that with zoomed DWI when the PSAD ratio is lower than 0.355. Low image quality, younger age, and lower PSAD contribute to csPCa undercategorized by mpMRI. Moreover, the use of zoomed DWI decreased undercategorized csPCa by improving PI-QUAL scores of MRI images.

Can artificial intelligence understand our emotions? Deep learning applications with face recognition

Abstract Objective The aim of this study is to evaluate the ability to detect emotions from human facial expressions via facial recognition technologies and analyze the effectiveness of deep learning models in this process. Method This research was conducted between 01.04 and 01.07.2024. The data of the study were taken from the open access site https://www.kaggle.com/datasets/msambare/fer2013 (Kaggle, 2024). Python 3.8 is used in this study. The FER-2013 (Facial Expression Recognition 2013) dataset is a comprehensive collection of facial images labeled with various emotions. The dataset contains 35,887 grayscale facial images. Each image has a size of 48 × 48 pixels. The dataset consists of images belonging to 7 emotion categories: anger, disgust, fear, happy, sad, confused, and neutral. Results In our experiments on the FER2013 dataset, we evaluated the performance of three different models: MobileNetV3-L, EfficientNetV2-L, and our proposed EfficientMobileNet. The evaluation criteria were based on sensitivity, specificity, accuracy, and F1 scores to assess the effectiveness of each model comprehensively. The EfficientMobileNet model outperformed MobileNetV3-L and EfficientNetV2-L in all measured performance metrics. EfficientMobileNet was the most successful model for predicting emotions, with an accuracy of 77.6%. Conclusion The impressive results obtained by EfficientMobileNet on the Fer2013 dataset show potential for wider application, especially in image classification tasks involving low-quality or small-scale images. This performance supports the idea of the potential for further improvements in neural network architecture and model efficiency and accuracy. Future work should focus on optimizing the model for more challenging datasets, studying the impact of different architectural adjustments, and investigating the scalability of EfficientMobileNet across various domains and applications.

DSpix2pix: A New Dual-Style Controlled Reconstruction Network for Remote Sensing Image Super-Resolution

Super-resolution reconstruction is a critical task in remote sensing image classification, and generative adversarial networks (GANs) have emerged as a dominant approach in this field. Traditional generative networks often produce low-quality images at resolutions like 256 × 256, and current research on single-image super-resolution typically focuses on resolution enhancement factors of two to four (2×–4×), which do not meet practical application demands. Building upon the framework of StyleGAN, this study introduces a dual-style controlled super-resolution reconstruction network referred to as DSpix2pix. It uses a fixed style vector (Style 1) from StyleGAN-v2, generated through its mapping network and applied to each layer in the generator. And an additional style vector (Style 2) is extracted from example images and injected into the decoder using AdIn, enhancing the balance of styles in the generated images. DSpix2pix is capable of generating high-quality, smoother, noise-reduced, and more realistic super-resolution remote sensing images at 512 × 512 and 1024 × 1024 resolutions. In terms of visual metrics such as RMSE, PSNR, SSIM, and LPIPS, it outperforms traditional super-resolution networks like SRGAN and UNIT, with RMSE consistently exceeding 10. The network excels in 2× and 4× super-resolution tasks, demonstrating potential for remote sensing image interpretation, and shows promising results in 8x super-resolution tasks.

Vision Transformers for Low-Quality Histopathological Images: A Case Study on Squamous Cell Carcinoma Margin Classification

Vision Transformers for Low-Quality Histopathological Images: A Case Study on Squamous Cell Carcinoma Margin Classification

Advanced Automated Model for Robust Bone Marrow Segmentation in Whole-body MRI.

To establish an advanced automated bone marrow (BM) segmentation model on whole-body (WB-)MRI in monoclonal plasma cell disorders (MPCD), and to demonstrate its robust performance on multicenter datasets with severe myeloma-related pathologies. The study cohort comprised multi-vendor, multi-protocol imaging data acquired with varying field strength across 8 different centers. In total, 210 WB-MRIs of 207 MPCD patients were included. An nnU-Net algorithm was established for segmenting the individual bone marrow spaces (BMS) of the spine, pelvis, humeri and femora (advanced segmentation model). For this task, 186 T1-weighted (T1w) WB-MRIs from center 1 were used in the training set. Test sets included 12 T1w WB-MRIs from center 2 (I) and 9 T1w WB-MRIs from centers 3-8 (II). Example cases were included to showcase segmentation performance on T1w WB-MRIs with extensive tumor load. The segmentation accuracy of the advanced segmentation model was compared to a prior established basic segmentation model by calculating Dice scores and using the Wilcoxon signed-rank test. The mean Dice score on the individual BMS was 0.89±0.13 (test set I) and 0.88±0.11 (test set II), significantly higher than the Dice scores of a prior basic model (p<0.05). Dice scores for the BMS of the individual bones ranged from 0.77 to 0.96 (test set I), and 0.81 to 0.95 (test set II). BM altered by myeloma-relevant pathologies, artifacts or low imaging quality was precisely segmented. The advanced model performed reliable, automated segmentations, even on heterogeneously acquired multicenter WB-MRIs with severe pathologies.

Neural Radiance Fields for High-Fidelity Soft Tissue Reconstruction in Endoscopy.

The advancement of neural radiance fields (NeRFs) has facilitated the high-quality 3D reconstruction of complex scenes. However, for most NeRFs, reconstructing 3D tissues from endoscopy images poses significant challenges due to the occlusion of soft tissue regions by invalid pixels, deformations in soft tissue, and poor image quality, which severely limits their application in endoscopic scenarios. To address the above issues, we propose a novel framework to reconstruct high-fidelity soft tissue scenes from low-quality endoscopic images. We first construct an EndoTissue dataset of soft tissue regions in endoscopic images and fine-tune the Segment Anything Model (SAM) based on EndoTissue to obtain a potent segmentation network. Given a sequence of monocular endoscopic images, this segmentation network can quickly obtain the tissue mask images. Additionally, we incorporate tissue masks into a dynamic scene reconstruction method called Tensor4D to effectively guide the reconstruction of 3D deformable soft tissues. Finally, we propose adopting the image enhancement model EDAU-Net to improve the quality of the rendered views. The experimental results show that our method can effectively focus on the soft tissue regions in the image, achieving higher fidelity in detail and geometric structural integrity in reconstruction compared to state-of-the-art algorithms. Feedback from the user study indicates high participant scores for our method.

A whole gamma imaging prototype for higher quantitative imaging of 89Zr-labeled antibodies in a tumor mouse model

Objective.Positron emission tomography (PET) has become an important clinical modality, but it is limited to imaging the annihilation radiation from positron-electron collisions. Recently, PET imaging with89Zr, which has a half-life of 3 d, has attracted much attention in immuno-PET to visualize immune cells and cancer cells by targeting specific antibodies on the cell surface. However,89Zr emits a single gamma ray at 909 keV four times more frequently than positrons, causing image quality (IQ) degradation in conventional PET. To overcome this drawback, use of such single gamma rays for imaging was previously proposed as whole gamma imaging (WGI). In WGI, a single gamma ray is detected by Compton imaging; by inserting a scatter detector ring inside the PET ring, WGI can realize both PET imaging and Compton imaging in one modality. A prototype for WGI was developed and Compton imaging of a mouse after intravenous administration of89Zr oxalate was demonstrated. However, the Compton imaging of the single gamma ray still presented a challenge due to its low IQ compared to PET.Approach.In this study, the scatter detector insert of the earlier WGI prototype was redesigned with the aim of improving Compton imaging performance. The new prototype produced WGI images by additive averaging of PET and Compton images after optimizing the ratio of each iteration in the image reconstruction. WGI IQ was then evaluated using the NEMA NU4 IQ phantom, and a tumor-burdened mouse was imaged with WGI up to 12 d after89Zr labeled antibody injection.Main results.Consequently, the Compton imaging performance was improved by lowering the angular resolution measure from 6.7 degrees to 6.4 degrees and the sensitivity from 0.11% to 0.18% compared to the previous prototype WGI. The phantom images with WGI showed a 15% reduction in noise and a 3% increase in contrast recovery under low-statistical conditions compared to images reconstructed by PET data alone.Significance. In-vivomouse imaging with the new prototype WGI was successfully performed. This successful imaging leads to the expectation that future whole-body WGI imaging will enable more sensitive and better quantitative89Zr antigen-antibody reaction imaging to be obtained.

Real-time human–computer interface based on eye gaze estimation from low-quality webcam images: integration of convolutional neural networks, calibration, and transfer learning

Abstract Eye gaze estimation represents a well-established research domain within computer vision. It has a wide range of practical applications in numerous fields, including human–computer interaction (HCI) for cursor control, health care, and virtual reality, enhancing its suitability for adoption throughout the scientific community. Different methods have been used for eye gaze estimation, such as model based, feature based, and appearance based. The appearance-based method is mainly used because it directly estimates an individual’s gaze direction from images/videos rather than depending on specific features or geometric models. This article developed an appearance-based, real-time generic eye gaze system for HCI to control the cursor through the eye using the convolutional neural network (CNN), calibration, and transfer learning. The study employed low-quality eye images captured from a conventional desktop webcam, enabling the proposed methodology to be implemented on any computer system equipped with a similar web camera without the need for supplementary hardware. Initially, the labeled dataset of both eyes is collected using the webcam. Then, a CNN model is trained by inputting left and right eye images to predict the gaze coordinate as output. We applied the calibration and transfer learning approach to the trained models to make a generic model for new users. In real-time use, the first step is calibration, where the user’s eye images are captured for various screen coordinates, and transfer learning is employed to fine-tune the pre-trained model according to the user’s eyes. Then, the fine-tuned model is used for eye gaze prediction to control the cursor. The system’s performance is evaluated using a test group of multiple users, and it demonstrated an average visual angle accuracy of 2.08 degrees before calibration, which notably improved to 1.81 degrees after the calibration process.

Machine learning of endoscopy images to identify, classify, and segment sinonasal masses.

We developed and assessed the performance of a machine learning model (MLM) to identify, classify, and segment sinonasal masses based on endoscopic appearance. A convolutional neural network-based model was constructed from nasal endoscopy images from patients evaluated at an otolaryngology center between 2013 and 2024. Images were classified into four groups: normal endoscopy, nasal polyps, benign, and malignant tumors. Polyps and tumors were confirmed with histopathological diagnosis. Images were annotated by an otolaryngologist and independently verified by two other otolaryngologists. We used high- and low-quality images to mirror real-world conditions. The models used for classification (EfficientNet-B2) and segmentation (nnUNet) were trained, validated, and tested at an 8:1:1 ratio. The performance accuracy was averaged across a 10-fold cross-validation assessment. Segmentation accuracy was assessed via Dice similarity coefficients. A total of 1242 images from 311 patients were used. The MLM was trained, validated, and tested on 663 normal, 276 polyps, 157 benign, and 146 malignant tumors images. Overall, the model performed at 84.1±4.3% accuracy in the validation set and 80.4±1.7% in the test set. The model correctly identified the presence of a sinonasal mass at 90.5±1.2% accuracy rate. The MLM accuracy performance rates were 86.2±1.0% for polyps and 84.1±1.8% for tumors. Benign and malignant tumor subclassification achieved 87.8±2.1% and 94.0±2.4% accuracy, respectively. Segmentation accuracies for polyps were 72.3% and 72.8% for tumors. An MLM for nasal endoscopy images can perform with moderate to high accuracy in identifying, classifying, and segmenting sinonasal masses. Performance in future iterations may improve with larger and more diverse training datasets.

Adaptive Evolutionary Optimization of Deep Learning Architectures for Focused Liver Ultrasound Image Segmentation.

Background: Liver ultrasound segmentation is challenging due to low image quality and variability. While deep learning (DL) models have been widely applied for medical segmentation, generic pre-configured models may not meet the specific requirements for targeted areas in liver ultrasound. Quantitative ultrasound (QUS) is emerging as a promising tool for liver fat measurement; however, accurately segmenting regions of interest within liver ultrasound images remains a challenge. Methods: We introduce a generalizable framework using an adaptive evolutionary genetic algorithm to optimize deep learning models, specifically U-Net, for focused liver segmentation. The algorithm simultaneously adjusts the depth (number of layers) and width (neurons per layer) of the network, dropout, and skip connections. Various architecture configurations are evaluated based on segmentation performance to find the optimal model for liver ultrasound images. Results: The model with a depth of 4 and filter sizes of [16, 64, 128, 256] achieved the highest mean adjusted Dice score of 0.921, outperforming the other configurations, using three-fold cross-validation with early stoppage. Conclusions: Adaptive evolutionary optimization enhances the deep learning architecture for liver ultrasound segmentation. Future work may extend this optimization to other imaging modalities and deep learning architectures.