Evaluation of medical X-ray images are traditionally carried out by visual inspection by medics, however some factors such as similar radiological findings, image quality, overlapping anatomical structure and subjectivity in interpretation make correct diagnosis difficult and may even delay treatments due to the need for a second medical opinion. The application of machine learning (ML) algorithms (MLA) to X-ray images can increase the accuracy and agility in disease diagnosis and assist doctors in decision-making. The objective of this pilot study is to test three deep learning (DL) architectures, including VGG16 and EfficientNet convolutional neural networks (CNN), evaluating their classification ability applied to chest X-rays, with and without lung pathologies and explore the use of Gradient-weighted Class Activation Mapping (Grad-CAM) as a tool for neural network interpretability. Images were taken from Kaggle datasets, and we used undersampling to prevent biases in training. The results demonstrate high accuracy in disease classification, particularly with EfficientNet. Grad-CAM was employed for interpretability of the trained CNN model, which produces a heatmap, highlighting the most important image regions for classification decisions. Despite promising results, analysis using Grad-CAM revealed room for improvement, prompting the need for collaboration with radiologists to evaluate the accuracy and clinical utility of our methodology. Future work may involve a hybrid approach, combining undersampling and oversampling techniques, using data augmentation or a larger variation of images, as well as the exploration of generative adversarial networks for superior results. Collaboration with the medical community is important for applying ML to medical imaging and analyzing tools for interpretability.

Computed tomography (CT) from full-angle data is of great importance to medical imaging and other areas. However, restrictions during data acquisition, such as measuring the target in a limited angular range, can lead to poorer quality results due to the increased difficulty of the image reconstruction problem. For challenging situations like this, traditional techniques may not be able to produce CT images with enough quality. In this study, we propose software for reconstruction of limited-angle CT images using a combination of deep image prior (DIP) and matching pursuit (MMP). Evaluation of our proposal was performed on the dataset provided by the Helsinki Tomographic Challenge 2022 (HTC 2022). Our results indicate that, by using MMP as a regularizer for the problem, the quality of DIP image reconstructions can be improved. Furthermore, we obtained images with a minimum Matthews correlation coefficient of 0.936 in five of seven difficulty levels of the HTC 2022, showing that it is a promising technique for this application.

The location of driver mechanisms in supra-ventricular tachyarrhythmias can be difficult to perform by human observers or in an automated manner, mainly when it comes to non-invasive methods. In this study, we investigated an automated methodology to locate two of the most common: Atrial Fibrillation (AF), driven by functional rotors; and atrial tachycardia (AT), driven by ectopic foci. In the proposed study, we used Farnebäck optical flow and calculated the divergence and curl of the generated vector field. First we validated the algorithm with signals from mathematical phantoms and then applied the method to AF and AT realistic in-silico models, comparing the results from the original simulated signals and the respective electrocardiographic imaging (ECGi), that is, after the originals went through forward solution followed by inverse solution. Results show that the present method is robust to noise, and outcomes are dependent of the predefined spatial resolution. In addition, the approach was able to locate both AT and AF mechanisms on the original atrial epicardial maps, and AF mechanism on the estimated atrial maps through the ECGi. However, AT mechanism location accuracy seems to be partially degraded by the ECGi.

Atrial Fibrillation (AF) is the most common supra-ventricular arrhythmia and has different underlying activation mechanism, including functional rotors (FR) and ectopic foci (EF). In this work we propose an approach for locating FR and EF in potential maps, which were tested with mathematical phantoms. 12 phantoms were created $\it(128x128$ array, 4 s, $Fs\ 500Hz)$ , simulating the motion of: FR (4 maps), EF (4 maps) and superpositions of these (4 maps). These were downsampled to different grids from $16x16$ to $8x6$ , simulating electrode acquisition. Noise (SNR from 2 to 60) was added. To locate the mechanism, the signals were filtered and interpolated. Farneback optical flow was applied to compute the motion vector field (MVF). The MVF was normalized and its temporal average was calculated. Finally, we computed curl and divergence, by using a $x$ and $y$ oriented $11 \times 11$ Sobel filter as $a$ estimation of the of the partial derivatives. The location of extrema in the curl and divergence maps were used for locating FR and EF respectively. Method robustness was tested by comparing the algorithm performance on different grid sizes and SNR. The mechanism was considered to be detected accurately if the position was within a normalized error of 5% from its respective phantom location. The results showed that our approach was able to locate both mechanisms, but revealed a dependency on spatial resolution.

Radiomics using rectal MRI radiomics has emerged as a promising approach in predicting pathological complete response. In this study, we present a typical pipeline of a radiomics analysis and review recent studies, exploring applications, development of radiomics methodologies and model construction in pCR prediction. Finally, we will offer our opinion about the future and discuss the next steps of rectal MRI radiomics for predicting pCR.

Pendelluft corresponds to the air flow from a nondependent lung region to a dependent one. Its occurrence during spontaneous breathing in mechanical ventilation is hard to detect in clinical routines and can lead to injuries. The aim of this study was to develop an automated methodology to diagnose Pendelluft from 2D + time Electrical Impedance Tomography sequences. Our images were acquired using an electrical impedance tomograph (1800 Enlight, Timpel, São Paulo, Brasil) from a normal pig (N) and a pig with the condition (P). Image analysis was performed in Matlab, divided into the following stages: (i) creation of motion vector field (MVF) using Horn and Schunck optical flow; (ii) decomposition of the MVF into a curl-free and divergence-free scalar potentials, D and R respectively, using the Discrete Helmholtz-Hodge Decomposition; (iii) the location of extrema in D and R were tracked over the image sequences N and P, and frequency maps created for the right and left lungs. The degree of similarity of these maps was represented by a parameter $$\phi $$ . The values of $$\phi $$ for P were zero for both lungs in the R and D analysis. In the D field analysis, pig N showed $$\phi $$ values of 6031 and 5407 for respective right and left lungs. The same analysis in the R field gave us a value of 4836 and 4538 for right and left lungs. These results show that our methodology is a possible candidate for automatic detection of Pendelluft, but studies from a large number of human subjects would be needed.

OBJECTIVES: To investigate whether quantitative textural features, extracted from pretreatment MRI, can predict sustained complete response to radiofrequency ablation (RFA) in patients with hepatocellular carcinoma (HCC). METHODS: In this IRB-approved study, patients were selected from a maintained six-year database of consecutive patients who underwent both pretreatment MRI imaging with a probable or definitive imaging diagnosis of HCC (LI-RADS 4 or 5) and loco-regional treatment with RFA. An experienced radiologist manually segmented the hepatic nodules in MRI arterial and equilibrium phases to obtain the volume of interest (VOI) for extraction of 107 quantitative textural features, including shape and first- and second-order features. Statistical analysis was performed to evaluate associations between textural features and complete response. RESULTS: The study consisted of 34 patients with 51 treated hepatic nodules. Sustained complete response was achieved by 6 patients (4 with single nodule and 2 with multiple nodules). Of the 107 features from the arterial and equilibrium phases, 20 (18%) and 25 (23%) achieved AUC >0.7, respectively. The three best performing features were found in the equilibrium phase: Dependence Non-Uniformity Normalized and Dependence Variance (both GLDM class, with AUC of 0.78 and 0.76, respectively) and Maximum Probability (GLCM class, AUC of 0.76). CONCLUSIONS: This pilot study demonstrates that a radiomic analysis of pre-treatment MRI might be useful in identifying patients with HCC who are most likely to have a sustained complete response to RFA. Second-order features (GLDM and GLCM) extracted from equilibrium phase obtained highest discriminatory performance.

In this study, we investigated the most relevant biomarkers for noninvasive classification and mechanism location in atrial tachycardia (AT), flutter (AFL) and fibrillation (AF). Biomarkers were calculated using noninvasive body surface (BSPM) dominant frequency and phase maps. We used 19 simulations of 567 to 64-lead BSPMs, from which were extracted 32 biomarkers. Biomarker ranking was performed with ANOVA, Kendall and Lasso techniques. The best four biomarkers were identified and used to classify the arrhythmias in all combinations, and the best two used for noninvasive driver localization. Arrhythmia classification accuracy was 94.74%. The feature combination which best distinguish AF from non-AF were mean filament displacement and mean OI, while those that best distinguish AFL from AT were mean and SD of SP distribution. There was good agreement across ranking techniques. Mechanism location accuracy was 78.95%, with the most important biomarkers being percentage SPs within each torso division, and SD of filament histogram cluster area. This study highlights that organization related features well identifies AF and spatial SP distribution discriminate AT from AFL and also it's localization.

Identifying the insertion zone of transcatheter heart valves can be time-consuming and suffers from variability and reproducibility problems. We present a deep leaning approach in CTA images to locate the midpoint of the insertion zone. A U-Net neural network is implemented to automatically segment the aortic valve on axial projection. The insertion zone midpoint is calculated based on the range of slices with the more concentrated area of activated pixels. We found a very low systematic error with a median computed error of 0.38mm and interquartile range of 0.15 – 0.75mm. The proposed model was shown to be a robust and powerful tool to automatically locate the insertion zone midpoint and we believe it will play a critical role on automated assessment of aortic stenosis.

The 2D Discrete Helmholtz Hodge Decomposition (DHHD) is a candidate for decomposition of cardiac motion vector field (MVFs) from modalities such as echocardiography, allowing analyses of rotational, radial and harmonic MVFs separately and the tracking of centers of flow phenomena. However, the literature contains differing opinions about how boundary conditions (BCs) should be applied. In the proposed study, we investigate two sets of BCs, found in the literature, using synthetic motion vector fields: (i) zero BCs for all nodes along the domain boundary; (ii) a single zero BC for the first node. We found that (i) produces a decomposition with Normalised Root Mean Square Error (NRMSE) below 1%, while (ii) produces NRMSE greater than 200% when the input field contains a harmonic component. Further, we demonstrate the decomposition of a synthetic motion field from the Extended Cardiac Torso Phantom v2 into radial, rotational and harmonic components.

Features of calcific leaflet distribution from CT valve images may allow prediction of hemodynamic disease severity in calcific degenerative aortic valve stenosis (DAS). The proposed study describes a signal processing scheme for selecting valve areas from maximum intensity projection valve images in a cohort of 52+43 patient images, diagnosed as having severe and moderate DAS. First the valve center and perimeter are approximated by a manually determined circle of radius r and center (x,y). The circle is then used to define eight masks, based on concentric circles of radius $\frac{r}{2},\frac{{r\sqrt 2 }}{2},\frac{{r\sqrt 3 }}{2}$ and r, each with center (x,y). The masks are used to define pixel regions within the valvular area, and statistical and textural descriptors are applied to each. Sensitivity/ specificity testing is performed with these descriptors, applied to the pixels within each mask, which show that disease severity is best predicted by using the smallest, most central mask and statistical features of skewness and kurtosis, providing area under the curve of 0.844 and 0.840 respectively. Our methodology was simple to implement and use, and provided good discriminatory power for disease severity. It also overcomes some difficulties in an earlier method, since our solution is scalable to variation in aortic valve size and tests a range ofstatistical and textural descriptors.

OBJECTIVES:Extract directional information related to left ventricular (LV) rotation and torsion from a 4D PET motion field using the Discrete Helmholtz Hodge Decomposition (DHHD).MATERIALS AND METHODS:Synthetic motion fields were created using superposition of rotational and radial field components and cardiac fields produced using optical flow from a control and patient image. These were decomposed into curl-free (CF) and divergence-free (DF) components using the DHHD.RESULTS:Synthetic radial components were present in the CF field and synthetic rotational components in the DF field, with each retaining its center position, direction of motion and diameter after decomposition. Direction of rotation at apex and base for the control field were in opposite directions during systole, reversing during diastole. The patient DF field had little overall rotation with several small rotators.CONCLUSIONS:The decomposition of the LV motion field into directional components could assist quantification of LV torsion, but further processing stages seem necessary.

The analysis of LV rotational motion could provide insights into myocardial dysfunction and predict the outcome of interventions, and this analysis could be performed more simply in separate rotational and radial components. In this study we present an automatic method for decomposing the cardiac motion field into radial and rotational components using the Discrete Helmholtz Hodge Decomposition (DHHD). The DHHD was applied to the following 3D motion fields (i) Synthetic complex motion fields, created by applying curl and gradient operators to Gaussian potentials, to determine numerical accuracy; (ii) Synthetic motion field from the 4D Extended Cardiac-Torso (XCAT) phantom (v2.0), to validate the use of the DHHD in decomposing cardiac motion fields. Decomposition error was found to decrease with increasing smoothness of the fields, while keeping motion field components small at the boundary of the motion field domain. The DHHD was seen to separate radial and rotational cardiac motion, allowing possible simplification ofmotion analysis.

Cardiac function can be assessed from displacement measurements in imaging modalities from nuclear medicine Using positron emission tomography (PET) image sequences with Rubidium-82, we propose and estimate the total Kinetic Energy Index (KEf) obtained from the velocity field, which was calculated using 3D optical flow(OF) methods applied over the temporal image sequence. However, it was found that the brightness of the image varied unexpectedly between frames, violating the constant brightness assumption of the OF method and causing large errors in estimating the velocity field. Therefore total brightness was equalized across image frames and the adjusted configuration tested with rest perfusion images acquired from individuals with normal (n=30) and low (n=33) cardiac function. For these images KEf was calculated as 0.5731±0.0899 and 0.3812±0.1146 for individuals with normal and low cardiac function respectively. The ability of KEf to properly classify patients into the two groups was tested with a ROC analysis, with area under the curve estimated as 0.906. To our knowledge this is the first time that KEf has been applied to PET images.

Automatic methods for quantifying velocity in cardiac images can be used to diagnose diseases, but the estimated velocity suffers from blurring in the region of motion discontinuities when using variational methods. In this paper, spatially varying filters were used to improve accuracy in these regions. These filters depend on: (i) image intensity and (ii) velocity estimates from previous iterations of the update equation. Methods were extended to voxel space, tested with synthetic images, and parameters optimized. The spatially varying filter which depends on velocity estimates was found to provide the greatest reduction in error measures, but the benefits were found to depend on careful choices of derivative scheme and other model parameters.