Relevant Image Features Research Articles

Predicting microvascular invasion in hepatocellular carcinoma: a deep learning model validated across hospitals

BackgroundThe accuracy of estimating microvascular invasion (MVI) preoperatively in hepatocellular carcinoma (HCC) by clinical observers is low. Most recent studies constructed MVI predictive models utilizing radiological and/or radiomics features extracted from computed tomography (CT) images. These methods, however, rely heavily on human experiences and require manual tumor contouring. We developed a deep learning-based framework for preoperative MVI prediction by using CT images of arterial phase (AP) with simple tumor labeling and without the need of manual feature extraction. The model was further validated on CT images that were originally scanned at multiple different hospitals.MethodsCT images of AP were acquired for 309 patients from China Medical University Hospital (CMUH). Images of 164 patients, who took their CT scanning at 54 different hospitals but were referred to CMUH, were also collected. Deep learning (ResNet-18) and machine learning (support vector machine) models were constructed with AP images and/or patients’ clinical factors (CFs), and their performance was compared systematically. All models were independently evaluated on two patient cohorts: validation set (within CMUH) and external set (other hospitals). Subsequently, explainability of the best model was visualized using gradient-weighted class activation map (Grad-CAM).ResultsThe ResNet-18 model built with AP images and patients’ clinical factors was superior than other models achieving a highest AUC of 0.845. When evaluating on the external set, the model produced an AUC of 0.777, approaching its performance on the validation set. Model interpretation with Grad-CAM revealed that MVI relevant imaging features on CT images were captured and learned by the ResNet-18 model.ConclusionsThis framework provide evidence showing the generalizability and robustness of ResNet-18 in predicting MVI using CT images of AP scanned at multiple different hospitals. Attention heatmaps obtained from model explainability further confirmed that ResNet-18 focused on imaging features on CT overlapping with the conditions used by radiologists to estimate MVI clinically.

Overview of Interventional Pulmonology for Radiologists.

Interventional pulmonology is a growing field specializing in minimally invasive procedures of the mediastinum, lungs, airways, and pleura. These procedures have both diagnostic and therapeutic indications and are performed for benign and malignant diseases. Endobronchial US has been combined with transbronchial needle aspiration to extend tissue sampling beyond the airways and into the lungs and mediastinum. Recent innovations extending the peripheral access of bronchoscopy include electromagnetic navigational bronchoscopy and thinner bronchoscopes. An important indication for therapeutic bronchoscopy is the relief of central airway obstruction, which may be severe and life threatening. Techniques for restoring patency of the central airways include mechanical debulking and multiple modalities for ablation, stent placement, and balloon bronchoplasty. Bronchoscopic lung volume reduction improves quality of life in certain patients with severe emphysema and is an important less invasive alternative to lung volume reduction surgery. Bronchial thermoplasty is likewise a nonpharmacologic treatment in patients with severe uncontrolled asthma. Many of these procedures have unique selection criteria that require precise evaluations at preprocedure imaging. Postprocedure imaging is also essential in determining outcome success and the presence of complications. Radiologists should be familiar with these procedures as well as the relevant imaging features in both planning and later surveillance. Evolving techniques that may become more widely available in the near future include robotic-assisted bronchoscopy, bronchoscopic transparenchymal nodule access, transbronchial cryobiopsy, ablation of early-stage cancers, and endobronchial intratumoral chemotherapy. An invited commentary by Wayne et al is available online. ©RSNA, 2021.

Macrovasculature and positron emission tomography (PET) standardized uptake value in patients with lung cancer.

To investigate the relationship between macrovasculature features and the standardized uptake value (SUV) of positron emission tomography (PET), which is a surrogate for the metabolic activity of a lung tumor. We retrospectively analyzed a cohort of 90 lung cancer patients who had both chest CT and PET-CT examinations before receiving cancer treatment. The SUVs in the medical reports were used. We quantified three macrovasculature features depicted on CT images (i.e., vessel number, vessel volume, and vessel tortuosity) and several tumor features (i.e., volume, maximum diameter, mean diameter, surface area, and density). Tumor size (e.g., volume) was used as a covariate to adjust for possible confounding factors. Backward stepwise multiple regression analysis was performed to develop a model for predicting PET SUV from the relevant image features. The Bonferroni correction was used for multiple comparisons. PET SUV was positively correlated with vessel volume (R=0.44, p<0.001) and vessel number (R=0.44, p<0.001) but not with vessel tortuosity (R=0.124, p>0.05). After adjusting for tumor size, PET SUV was significantly correlated with vessel tortuosity (R=0.299, p=0.004) and vessel number (R=0.224, p=0.035), but only marginally correlated with vessel volume (R=0.187, p=0.079). The multiple regression model showed a performance with an R-Squared of 0.391 and an adjusted R-Squared of 0.355 (p<0.001). Our investigations demonstrate the potential relationship between macrovasculature and PET SUV and suggest the possibility of inferring the metabolic activity of a lung tumor from chest CT images.

Primary renal sarcomas: imaging features and discrimination from non-sarcoma renal tumors

ObjectivesTo assess imaging features of primary renal sarcomas in order to better discriminate them from non-sarcoma renal tumors.MethodsAdult patients diagnosed with renal sarcomas from 1995 to 2018 were included from 11 European tertiary referral centers (Germany, Belgium, Turkey). Renal sarcomas were 1:4 compared to patients with non-sarcoma renal tumors. CT/MRI findings were assessed using 21 predefined imaging features. A random forest model was trained to predict “renal sarcoma vs. non-sarcoma renal tumors” based on demographics and imaging features.Resultsn = 34 renal sarcomas were included and compared to n = 136 non-sarcoma renal tumors. Renal sarcomas manifested in younger patients (median 55 vs. 67 years, p < 0.01) and were more complex (high RENAL score complexity 79.4% vs. 25.7%, p < 0.01). Renal sarcomas were larger (median diameter 108 vs. 43 mm, p < 0.01) with irregular shape and ill-defined margins, and more frequently demonstrated invasion of the renal vein or inferior vena cava, tumor necrosis, direct invasion of adjacent organs, and contact to renal artery or vein, compared to non-sarcoma renal tumors (p < 0.05, each). The random forest algorithm yielded a median AUC = 93.8% to predict renal sarcoma histology, with sensitivity, specificity, and positive predictive value of 90.4%, 76.5%, and 93.9%, respectively. Tumor diameter and RENAL score were the most relevant imaging features for renal sarcoma identification.ConclusionRenal sarcomas are rare tumors commonly manifesting as large masses in young patients. A random forest model using demographics and imaging features shows good diagnostic accuracy for discrimination of renal sarcomas from non-sarcoma renal tumors, which might aid in clinical decision-making.Key Points• Renal sarcomas commonly manifest in younger patients as large, complex renal masses.• Compared to non-sarcoma renal tumors, renal sarcomas more frequently demonstrated invasion of the renal vein or inferior vena cava, tumor necrosis, direct invasion of adjacent organs, and contact to renal artery or vein.• Using demographics and standardized imaging features, a random forest showed excellent diagnostic performance for discrimination of sarcoma vs. non-sarcoma renal tumors (AUC = 93.8%, sensitivity = 90.4%, specificity = 76.5%, and PPV = 93.9%).

A new hardware Trojan detection technique using deep convolutional neural network

The involvement of external vendors in semiconductor industries increases the chance of hardware Trojan (HT) insertion in different phases of the integrated circuit (IC) design. Recently, several partial reverse engineering (RE) based HT detection techniques are reported, which attempt to reduce the time and complexity involved in the full RE process by applying machine learning or image processing techniques in IC images. However, these techniques fail to extract the relevant image features, not robust to image variations, complicated, less generalizable, and possess a low detection rate. Therefore, to overcome the above limitations, this paper proposes a new partial RE based HT detection technique that detects Trojans from IC layout images using Deep Convolutional Neural Network (DCNN). The proposed DCNN model consists of stacking several convolutional and pooling layers. It layer-wise extracts and selects the most relevant and robust features automatically from the IC images and eliminates the need to apply the feature extraction algorithm separately. To prevent the over-training of the DCNN model, a new stopping condition method and two new metrics, namely Accuracy difference measure (ADM) and Loss difference measure (LDM), are proposed that halts the training only when the performance of our model genuinely drops. Further, to combat the issue of process variations and fabrication noise generated during the RE process, we include noisy images with varying parameters in the training process of the model. We also apply the data augmentation and regularization techniques in the model to address the issues of underfitting and overfitting. Experimental evaluation shows that the proposed technique provides 99% and 97.4% accuracy on Trust-Hub and synthetic ISCAS dataset, respectively, which is on-an-average 15.83% and 21.69% higher than the existing partial RE based techniques.

Physiologically decreased F-18 fluorodeoxyglucose uptake in the lower vertebrae associated with daily drinking habit in Japanese men with alcohol flushing reaction.

Alcohol flushing reaction (AFR) is known as one of the risks for esophageal squamous cell cancer, and scientists have been elucidating this issue. However, little attention has been given to relevant imaging features. This study aims to investigate whether physiological 18F-fluorodeoxyglucose (FDG) uptake patterns in vertebrae are associated with drinking habits or AFR. Japanese male patients who underwent FDG positron emission computed tomography for evaluation of their known or suspected malignancies or inflammatory diseases were asked about their drinking habits and AFR. Altogether, 192 patients, 139 every-day drinkers and 53 non-drinkers were evaluated. Comparing the FDG uptake between that in the thoracic region and that in the lumbar region, vertebral uptake was visually classified into four patterns: Ld, dominant in lumbar region; TL, almost equal in both regions; BL, slightly higher in thoracic region (borderline pattern); Td, dominant in thoracic region. The uptake patterns were evaluated according to drinking habit (every-day drinker or non-drinker), AFR (flusher or non-flusher), and the combination of these two factors (habit/reaction: every-day drinker/flusher, every-day drinker/non-flusher, non-drinker/flusher, or non-drinker/non-flusher). There were 95 flushers (51 every-day drinkers and 44 non-drinkers) and 97 non-flushers (88 every-day drinkers and 9 non-drinkers). Ld, TL, BL, and Td patterns were observed in 0, 109 (56.8%), 31 (16.1%), and 52 (27.1%) patients, respectively. Td and BL patterns were more frequently observed in every-day drinkers compared with non-drinkers (p=0.0467). Though the uptake patterns did not differ between flushers and non-flushers (p=0.116), the Td pattern was more frequently observed in every-day drinkers/flushers (51%) compared with every-day drinkers/non-flushers (20.5%), non-drinkers/flushers (13.6%), and non-drinkers/non-flushers (22.2%) (p=0.0014). The Td pattern was observed in patients with various diseases, with higher frequency in esophageal cancer, head and neck cancer, and lung cancer compared with other diseases. In conclusion, drinking habits and AFR were related to the vertebral uptake pattern with decreased uptake in the lumbar region in Japanese male patients.

SAUN: Stack attention U-Net for left ventricle segmentation from cardiac cine magnetic resonance imaging.

PurposeQuantification of left ventricular (LV) volume, ejection fraction and myocardial mass from multi‐slice multi‐phase cine MRI requires accurate segmentation of the LV in many images. We propose a stack attention‐based convolutional neural network (CNN) approach for fully automatic segmentation from short‐axis cine MR images.MethodsTo extract the relevant spatiotemporal image features, we introduce two kinds of stack methods, spatial stack model and temporal stack model, combining the target image with its neighboring images as the input of a CNN. A stack attention mechanism is proposed to weigh neighboring image slices in order to extract the relevant features using the target image as a guide. Based on stack attention and standard U‐Net, a novel Stack Attention U‐Net (SAUN) is proposed and trained to perform the semantic segmentation task. A loss function combining cross‐entropy and Dice is used to train SAUN. The performance of the proposed method was evaluated on an internal and a public dataset using technical metrics including Dice, Hausdorff distance (HD), and mean contour distance (MCD), as well as clinical parameters, including left ventricular ejection fraction (LVEF) and myocardial mass (LVM). In addition, the results of SAUN were compared to previously presented CNN methods, including U‐Net and SegNet.ResultsThe spatial stack attention model resulted in better segmentation results than the temporal stack model. On the internal dataset comprising of 167 post‐myocardial infarction patients and 57 healthy volunteers, our method achieved a mean Dice of 0.91, HD of 3.37 mm, and MCD of 1.08 mm. Evaluation on the publicly available ACDC dataset demonstrated good generalization performance, yielding a Dice of 0.92, HD of 9.4 mm, and MCD of 0.74 mm on end‐diastolic images, and a Dice of 0.89, HD of 7.1 mm and MCD of 1.03 mm on end‐systolic images. The Pearson correlation coefficient of LVEF and LVM between automatically and manually derived results were higher than 0.98 in both datasets.ConclusionWe developed a CNN with a stack attention mechanism to automatically segment the LV chamber and myocardium from the multi‐slice short‐axis cine MRI. The experimental results demonstrate that the proposed approach exceeds existing state‐of‐the‐art segmentation methods and verify its potential clinical applicability.

Development and content validity evaluation of a candidate instrument to assess image quality in digital mammography: A mixed-method study

PurposeTo develop a candidate instrument to assess image quality in digital mammography, by identifying clinically relevant features in images that are affected by lower image quality. MethodsInterviews with fifteen expert breast radiologists from five countries were conducted and analysed by using adapted directed content analysis. During these interviews, 45 mammographic cases, containing 44 lesions (30 cancers, 14 benign findings), and 5 normal cases, were shown with varying image quality. The interviews were performed to identify the structures from breast tissue and lesions relevant for image interpretation, and to investigate how image quality affected the visibility of those structures. The interview findings were used to develop tentative items, which were evaluated in terms of wording, understandability, and ambiguity with expert breast radiologists. The relevance of the tentative items was evaluated using the content validity index (CVI) and modified kappa index (k*). ResultsTwelve content areas, representing the content of image quality in digital mammography, emerged from the interviews and were converted into 29 tentative items. Fourteen of these items demonstrated excellent CVI ≥ 0.78 (k* > 0.74), one showed good CVI < 0.78 (0.60 ≤ k* ≤ 0.74), while fourteen were of fair or poor CVI < 0.78 (k* ≤ 0.59). In total, nine items were deleted and five were revised or combined resulting in 18 items. ConclusionsBy following a mixed-method methodology, a candidate instrument was developed that may be used to characterise the clinically-relevant impact that image quality variations can have on digital mammography.

Preoperative Prediction Nomogram Based on Integrated Profiling for Glioblastoma Multiforme in Glioma Patients.

Introduction: Traditional classification that divided gliomas into glioblastoma multiformes (GBM) and lower grade gliomas (LGG) based on pathological morphology has been challenged over the past decade by improvements in molecular stratification, however, the reproducibility and diagnostic accuracy of glioma classification still remains poor. This study aimed to establish and validate a novel nomogram for the preoperative diagnosis of GBM by using integrated data combined with feasible baseline characteristics and preoperative tests.Material and method: The models were established in a primary cohort that included 259 glioma patients who had undergone surgical resection and were pathologically diagnosed from March 2014 to May 2016 in the First Affiliated Hospital of Xi'an Jiaotong University. The preoperative data were used to construct three models by the best subset regression, the forward stepwise regression, and the least absolute shrinkage and selection operator, and to furthermore establish the nomogram among those models. The assessment of nomogram was carried out by the discrimination and calibration in internal cohorts and external cohorts.Results and discussion: Out of all three models, model 2 contained eight clinical-related variables, which exhibited the minimum Akaike Information Criterion (173.71) and maximum concordance index (0.894). Compared with the other two models, the integrated discrimination index for model 2 was significantly improved, indicating that the nomogram obtained from model 2 was the most appropriate model. Likewise, the nomogram showed great calibration and significant clinical benefit according to calibration curves and the decision curve analysis.Conclusion: In conclusion, our study showed a novel preoperative model that incorporated clinically relevant variables and imaging features with laboratory data that could be used for preoperative prediction in glioma patients, thus providing more reliable evidence for surgical decision-making.

ERT Image Evaluation Based on Sparse Representation Algorithm

Most of the existing image quality evaluation is to extract image features, and then use support vector regression, sparse representation and other algorithms to evaluate the image quality. In this paper, an adaptive sub-dictionary image evaluation algorithm based on sparse representation is used. By extracting relevant image features such as gradient features and color features, a complete dictionary is first constructed using the features, and then a sparse representation method based on the sub-dictionary is used to obtain the sparse corresponding to the ERT image. Coefficient, the final image quality score is obtained by the corresponding formula.

Deep Learning for Whole-Slide Tissue Histopathology Classification: A Comparative Study in the Identification of Dysplastic and Non-Dysplastic Barrett's Esophagus.

The gold standard of histopathology for the diagnosis of Barrett’s esophagus (BE) is hindered by inter-observer variability among gastrointestinal pathologists. Deep learning-based approaches have shown promising results in the analysis of whole-slide tissue histopathology images (WSIs). We performed a comparative study to elucidate the characteristics and behaviors of different deep learning-based feature representation approaches for the WSI-based diagnosis of diseased esophageal architectures, namely, dysplastic and non-dysplastic BE. The results showed that if appropriate settings are chosen, the unsupervised feature representation approach is capable of extracting more relevant image features from WSIs to classify and locate the precursors of esophageal cancer compared to weakly supervised and fully supervised approaches.

Histopathological imaging-based cancer heterogeneity analysis via penalized fusion with model averaging.

Heterogeneity is a hallmark of cancer. For various cancer outcomes/phenotypes, supervised heterogeneity analysis has been conducted, leading to a deeper understanding of disease biology and customized clinical decisions. In the literature, such analysis has been oftentimes based on demographic, clinical, and omics measurements. Recent studies have shown that high-dimensional histopathological imaging features contain valuable information on cancer outcomes. However, comparatively, heterogeneity analysis based on imaging features has been very limited. In this article, we conduct supervised cancer heterogeneity analysis using histopathological imaging features. The penalized fusion technique, which has notable advantages-such as greater flexibility-over the finite mixture modeling and other techniques, is adopted. A sparse penalization is further imposed to accommodate high dimensionality and select relevant imaging features. To improve computational feasibility and generate more reliable estimation, we employ model averaging. Computational and statistical properties of the proposed approach are carefully investigated. Simulation demonstrates its favorable performance. The analysis of The Cancer Genome Atlas (TCGA) data may provide a new way of defining/examining breast cancerheterogeneity.

Effects of hydrogen and primary air in a commercial partially-premixed atmospheric gas burner by means of optical and supervised machine learning techniques

In order to ascertain the effects of the hydrogen addition and the primary air-fuel ratio on burner performance and emissions, we conduct tests on a commercial atmospheric gas burner using pure methane and a blend of hydrogen/methane. Relevant statistical image features are extracted from a UV–VIS camera equipped with narrow-band optical filters. Radical image results agrees with spectrometric data, showing the relevance of the OH∗ intensity radiation coming from the outer non-premixed zone. The double-cone flame structure is evident, showing a growing secondary non-premixed cone as the primary air-fuel ratio is decreased. In addition, the direct relationship found between flame radical imaging features and NOx emissions has been used to develop a predictive model by integrating classification techniques and neural networks. The research confirms UV–VIS chemiluminescence imaging techniques as powerful tools aimed at combustion monitoring, with huge prospects of being integrated within advanced emission control techniques for commercial burners.

Differentiating rheumatoid and psoriatic arthritis: a systematic analysis of high-resolution magnetic resonance imaging features\u2014preliminary findings

BackgroundBecause of overlapping phenotypical presentations, the diagnostic differentiation of rheumatoid arthritis (RA) and psoriatic arthritis (PsA) remains challenging. Thus, this study aimed to examine the diagnostic value of distinct imaging features obtained by high-resolution 3-T MRI for the diagnostic differentiation.Materials and methodsSeventeen patients with PsA and 28 patients with RA were imaged at high resolution using 3-T MRI scanners and a dedicated 16-channel hand coil. All images were analyzed according to the outcome measures in rheumatology clinical trials’ (OMERACT) RAMRIS (Rheumatoid Arthritis Magnetic Resonance Imaging Score) and PsAMRIS (Psoriatic Arthritis Magnetic Resonance Imaging Score) for the presence and intensity of synovitis, flexor tenosynovitis, bone edema, bone erosion, periarticular inflammation, bone proliferation, and joint space narrowing. Next, odds ratios (OR) were calculated to determine the strength of the associations between these imaging features, demographic characteristics, and the outcome RA vs. PsA.ResultsPsA could be differentiated from RA by extracapsular inflammatory changes (PsAMRIS sub-score “periarticular inflammation”), with low odds for the presence of RA (OR of 0.06, p < 0.01) at all metacarpophalangeal (MCP) joints. A prediction model informed by the items that were strongest associated with the presence of RA or PsA demonstrated excellent differentiating capability with an area under the curve of 98.1%.ConclusionHigh-resolution imaging is beneficial for the identification of relevant imaging features that may assist the clinical differentiation of inflammatory conditions of the hand. At the MCP level, extracapsular inflammatory changes were strongly associated with PsA and may consequently allow the imaging differentiation of PsA and RA.

Efficient deep discriminant embedding: Application to face beauty prediction and classification

Inspired by deep learning architectures, we introduce a multi-layer local discriminant embedding algorithm that integrates feature selection as a main step to capture the most relevant and discriminant features of an input face image or face descriptor. The proposed framework allows to transform any linear method to a deep variant via a cascaded feature extraction and selection architecture able to convert weak and noisy descriptors to strong ones. As a case study, the local discriminant embedding (LDE) projection is adopted as a linear feature extraction method. The resulting framework can be considered as an efficient deep discriminant embedding technique. To validate this framework, we have considered two different computer vision problems: face beauty prediction which involves both classification and regression tasks, and face recognition which is a classical classification problem. Experiments conducted on different public benchmark databases show that this approach enhances the performance of the LDE algorithm and provides a discriminating strategy to solve the dimensionality reduction problem. For face beauty regression, our proposed framework achieved on average an improvement of about 5% and 7% with respect to two other configurations where only VGG-face and VGG-face followed by LDE have been considered. For face beauty classification, the proposed algorithm outperformed many classical manifold learning techniques reaching in some databases improvements of about 10%.

Concept attribution: Explaining CNN decisions to physicians

Deep learning explainability is often reached by gradient-based approaches that attribute the network output to perturbations of the input pixels. However, the relevance of input pixels may be difficult to relate to relevant image features in some applications, e.g. diagnostic measures in medical imaging. The framework described in this paper shifts the attribution focus from pixel values to user-defined concepts. By checking if certain diagnostic measures are present in the learned representations, experts can explain and entrust the network output. Being post-hoc, our method does not alter the network training and can be easily plugged into the latest state-of-the-art convolutional networks. This paper presents the main components of the framework for attribution to concepts, in addition to the introduction of a spatial pooling operation on top of the feature maps to obtain a solid interpretability analysis. Furthermore, regularized regression is analyzed as a solution to the regression overfitting in high-dimensionality latent spaces. The versatility of the proposed approach is shown by experiments on two medical applications, namely histopathology and retinopathy, and on one non-medical task, the task of handwritten digit classification. The obtained explanations are in line with clinicians’ guidelines and complementary to widely used visualization tools such as saliency maps.

Imaging in trauma of the facial skeleton and soft tissues of the neck.

Trauma to the face and neck is a frequent reason for emergency department attendance. Imaging is invaluable in the characterisation of such injuries, enabling delineation of fracture patterns as well as identification of vascular and other soft tissue injuries. It may also be used to prevent long-term mortality and morbidity and provide a roadmap for surgical intervention so that form and function may be restored. This article gives a pictorial review of the imaging of craniofacial trauma, stratified according to the thirds of the face, followed by a review of blunt and penetrating trauma of the neck. It discusses appropriate imaging modalities for each trauma category, describes major patterns of craniofacial trauma on cross-sectional imaging and identifies clinically relevant imaging features that should trigger subspecialist review or be of relevance to pre-surgical planning. It starts with the upper third comprising frontal sinus fractures before describing the component fractures of the middle third (including nasal, zygomaticomaxillary and orbital fractures) and then focusing on the lower third (specifically mandibular and dentoalveolar fractures). The article concludes with a review of soft tissue injuries of the neck, particularly penetrating, blunt and laryngeal trauma.

Cloud Identification from All-sky Camera Data with Machine Learning

Abstract Most ground-based observatories are equipped with wide-angle all-sky cameras to monitor the night sky conditions. Such camera systems can be used to provide an early warning of incoming clouds that can pose a danger to the telescope equipment through precipitation, as well as for sky quality monitoring. We investigate the use of different machine-learning approaches for automating the identification of mostly opaque clouds in all-sky camera data as a cloud warning system. In a deep-learning approach, we train a residual neural network (ResNet) on pre-labeled camera images. Our second approach extracts relevant and localized image features from camera images and uses these data to train a gradient-boosted tree-based model (lightGBM). We train both model approaches on a set of roughly 2000 images taken by the all-sky camera located at Lowell Observatory’s Discovery Channel Telescope, in which the presence of clouds has been labeled manually. The ResNet approach reaches an accuracy of 85% in detecting clouds in a given region of an image, but requires a significant amount of computing resources. Our lightGBM approach achieves an accuracy of 95% with a training sample of ∼1000 images and rather modest computing resources. Based on different performance metrics, we recommend the latter feature-based approach for automated cloud detection. Code that was built for this work is available online.

Towards expert gaze modeling and recognition of a user’s attention in realtime

Abstract One of the appealing areas of expertise research is devoted to measuring the effectiveness of training programs for novices. With recent progress in eye tracking, gaze-based interaction systems recognize a user’s attention and can direct it accordingly. Moreover, dynamic visualization of an expert gaze model facilitates novice training by guiding the gaze to relevant areas. In addition, the system should be aware of realtime attention to remove an overlay that could occlude relevant information. We use an implementation of subtle gaze direction (SGD) and the simplified scanpath of a dentist to train naive participants in finding anomalies in dental radiographs. We were able to effectively direct user gaze to relevant image features without occluding the area when attention was recognized. Additionally, participants reported that the intervention was helpful for image inspection. The results of the model intervention show minimal improvements in anomaly detection, which is expected of naive subjects. We advocate that the system has the potential to be highly effective for advanced students and trainees with a certain foundation of conceptual knowledge.

A Classification Algorithm by Combination of Feature Decomposition and Kernel Discriminant Analysis (KDA) for Automatic MR Brain Image Classification and AD Diagnosis.

Magnetic resonance (MR) imaging is a widely used imaging modality for detection of brain anatomical variations caused by brain diseases such as Alzheimer's disease (AD) and mild cognitive impairment (MCI). AD considered as an irreversible neurodegenerative disorder with progressive memory impairment moreover cognitive functions, while MCI would be considered as a transitional phase amongst age-related cognitive weakening. Numerous machine learning approaches have been examined, aiming at AD computer-aided diagnosis through employing MR image analysis. Conversely, MR brain image changes could be caused by different effects such as aging and dementia. It is still a challenging difficulty to extract the relevant imaging features and classify the subjects of different groups. This paper would propose an automatic classification technique based on feature decomposition and kernel discriminant analysis (KDA) for classifications of progressive MCI (pMCI) vs. normal control (NC), AD vs. NC, and pMCI vs. stable MCI (sMCI). Feature decomposition would be based on dictionary learning, which is used for separation of class-specific components from the non-class-specific components in the features, while KDA would be applied for mapping original nonlinearly separable feature space to the separable features that are linear. The proposed technique would be evaluated by employing T1-weighted MR brain images from 830 subjects comprising 198 AD patients, 167 pMCI, 236 sMCI, and 229 NC from the Alzheimer's disease neuroimaging initiative (ADNI) dataset. Experimental results demonstrate that classification accuracy (ACC) of 90.41%, 84.29%, and 65.94% can be achieved for classification of AD vs. NC, pMCI vs. NC, and pMCI vs. sMCI, respectively, indicating the promising performance of the proposed method.