Large Set Of Images Research Articles

A Comparative Analysis of Content based Image Retrieval

With the development of multimedia technology, the rapid increasing usage of large image database becomes possible. To carry out its management and retrieval, Content-Based Image Retrieval (CBIR) is an effective method. It will be very difficult to manage this database of images stored at the remote servers. The right tool will be required which can process these images for different operations. These operations include searching etc. It will be difficult to classify the images into groups and then search each class for providing the image as the information against the user request query. The content based image retrieval is the most suitable way to identify the image from the large repository. It will search the image from the large set of images based on contents rather than the image name. It will be having less time to search the image from the large repository when the image is retrieved using content based. In the current research the hybrid approach for content based image retrieval is performed. This proposed procedure will be in the first step perform the classification of the image into multiple classes. The classes are prepared based on the attributes values.

A nonlocal deep image prior model to restore optical coherence tomographic images from gamma distributed speckle noise

Optical Coherence Tomography (OCT) is often employed to observe the retinal layers in the human eyes. The retinal scans are susceptible to artefacts such as head movements or eye blinks. Along with this, the quality of the images is degraded by speckle noise caused due to the constructive and destructive interference of the waves used for capturing data. Recently, image restoration techniques have geared up in terms of quality with the exertion of deep learning. Despeckling using deep learning, in general, necessitates a large set of training images. On the contrary, deep image prior is a novel model that performs denoising operations using a single training image, based on a prior assumption about the noise distribution. This paper extends the concept of the deep image prior towards non-local restoration for speckle noise assuming that the speckle follows Gamma distribution. Such a framework can be incorporated to enhance the OCT images. The proposed framework is assessed qualitatively with visual comparisons and quantitatively using statistical measures like PSNR, CNR and ENL. Comparative studies confirm that the proposed method outperforms the existing methods in restoring speckled input images.

On the Feasibility of Fast Fourier Transform Separability Property for Distributed Image Processing

Given the high algorithmic complexity of applied-to-images Fast Fourier Transforms (FFT), computational-resource-usage efficiency has been a challenge in several engineering fields. Accelerator devices such as Graphics Processing Units are very attractive solutions that greatly improve processing times. However, when the number of images to be processed is large, having a limited amount of memory is a serious problem. This can be faced by using more accelerators or using higher-capability accelerators, which implies higher costs. The separability property is a resource in hardware approaches that is frequently used to divide the two-dimensional FFT work into several one-dimensional FFTs, which can be simultaneously processed by several computing units. Then, a feasible alternative to address this problem is distributed computing through an Apache Spark cluster. However, determining the separability-property feasibility in distributed systems, when migrating from hardware implementations, is not evident. For this reason, in this paper a comparative study is presented between distributed versions of two-dimensional FFTs using the separability property to determine the suitable way to process large image sets using both Spark RRDs and DataFrame APIs.

Critical information thresholds underlying generic and familiar face categorisation at the same face encounter

Seeing a face in the real world provokes a host of automatic categorisations related to sex, emotion, identity, and more. Such individual facets of human face recognition have been extensively examined using overt categorisation judgements, yet their relative informational dependencies during the same face encounter are comparatively unknown. Here we used EEG to assess how increasing access to sensory input governs two ecologically relevant brain functions elicited by seeing a face: Distinguishing faces and nonfaces, and recognising people we know. Observers viewed a large set of natural images that progressively increased in either image duration (experiment 1) or spatial frequency content (experiment 2). We show that in the absence of an explicit categorisation task, the human brain requires less sensory input to categorise a stimulus as a face than it does to recognise whether that face is familiar. Moreover, where sensory thresholds for distinguishing faces/nonfaces were remarkably consistent across observers, there was high inter-individual variability in the lower informational bound for familiar face recognition, underscoring the neurofunctional distinction between these categorisation functions. By i) indexing a form of face recognition that goes beyond simple low-level differences between categories, and ii) tapping multiple recognition functions elicited by the same face encounters, the information minima we report bear high relevance to real-world face encounters, where the same stimulus is categorised along multiple dimensions at once. Thus, our finding of lower informational requirements for generic vs. familiar face recognition constitutes some of the strongest evidence to date for the intuitive notion that sensory input demands should be lower for recognising face category than face identity.

Cultural Analytics of Large Datasets from Flickr

Deluge became a metaphor to describe the amount of information to which we are subjected, and very often we feel we are drowning while our access to information is rising. Devising mechanisms for exploring massive image sets according to perceptual attributes is still a challenge, even more when dealing with user-generated social media content. Such images tend to be heterogenous, and using metadata-only can be misleading. This paper describes a set of tools designed to analyze large sets of user-created art related images using image features describing color, texture, composition and orientation. The proposed pipeline permits to discriminate Flickr groups in terms of feature vectors and clustering parameters. The algorithms are general enough to be applied to other domains in which the main question is about the variability of the images.

Local Style Preservation in Improved GAN-Driven Synthetic Image Generation for Endoscopic Tool Segmentation.

Accurate semantic image segmentation from medical imaging can enable intelligent vision-based assistance in robot-assisted minimally invasive surgery. The human body and surgical procedures are highly dynamic. While machine-vision presents a promising approach, sufficiently large training image sets for robust performance are either costly or unavailable. This work examines three novel generative adversarial network (GAN) methods of providing usable synthetic tool images using only surgical background images and a few real tool images. The best of these three novel approaches generates realistic tool textures while preserving local background content by incorporating both a style preservation and a content loss component into the proposed multi-level loss function. The approach is quantitatively evaluated, and results suggest that the synthetically generated training tool images enhance UNet tool segmentation performance. More specifically, with a random set of 100 cadaver and live endoscopic images from the University of Washington Sinus Dataset, the UNet trained with synthetically generated images using the presented method resulted in 35.7% and 30.6% improvement over using purely real images in mean Dice coefficient and Intersection over Union scores, respectively. This study is promising towards the use of more widely available and routine screening endoscopy to preoperatively generate synthetic training tool images for intraoperative UNet tool segmentation.

Automatic Diagnosis of Spinal Disorders on Radiographic Images: Leveraging Existing Unstructured Datasets With Natural Language Processing.

Retrospective study. Huge amounts of images and medical reports are being generated in radiology departments. While these datasets can potentially be employed to train artificial intelligence tools to detect findings on radiological images, the unstructured nature of the reports limits the accessibility of information. In this study, we tested if natural language processing (NLP) can be useful to generate training data for deep learning models analyzing planar radiographs of the lumbar spine. NLP classifiers based on the Bidirectional Encoder Representations from Transformers (BERT) model able to extract structured information from radiological reports were developed and used to generate annotations for a large set of radiographic images of the lumbar spine (N = 10 287). Deep learning (ResNet-18) models aimed at detecting radiological findings directly from the images were then trained and tested on a set of 204 human-annotated images. The NLP models had accuracies between 0.88 and 0.98 and specificities between 0.84 and 0.99; 7 out of 12 radiological findings had sensitivity >0.90. The ResNet-18 models showed performances dependent on the specific radiological findings with sensitivities and specificities between 0.53 and 0.93. NLP generates valuable data to train deep learning models able to detect radiological findings in spine images. Despite the noisy nature of reports and NLP predictions, this approach effectively mitigates the difficulties associated with the manual annotation of large quantities of data and opens the way to the era of big data for artificial intelligence in musculoskeletal radiology.

A Comparative Study of Handwritten Character Recognition by using Image Processing and Neural Network Techniques

This study aims to analyze the effects of noise, image filtering, and edge detection techniques in the preprocessing phase of character recognition by using a large set of character images exported from MNIST database trained with various sizes of neural networks. Canny edge detection algorithm was deployed to smooth the edges of the images while the Sobel edge detection algorithm was used to detect the edges of the images. Skeletonization algorithm was applied to re-shape the structural shapes. In the context of the image filtering, the Laplacian filter was utilized to enhance the images and High pass filtering was used to highlight the fine details in blurred images. Gaussian noise, image noise with Gaussian intensity, function in Matlab with the probability density function P was deployed on character images of MINST. Pattern recognition neural networks are widely used in optical character recognition. Feedforward neural networks are deployed in this study. A comprehensive analysis of the above-mentioned image processing techniques is included during character recognition. Improved accuracy is observed with character recognition during the prediction phase of the neural networks. A sample of unknown characters is tested with the application of High pass filtering + feedforward neural network and 89%, the highest, average output prediction accuracy was obtained. Other prediction accuracies were also tabulated for the reader’s attention.

PoreAnalyzer—An Open-Source Framework for the Analysis and Classification of Defects in Additive Manufacturing

Additive manufacturing processes offer high geometric flexibility and allow the use of new alloy concepts due to high cooling rates. For each new material, parameter studies have to be performed to find process parameters that minimize microstructural defects such as pores or cracks. In this paper, we present a system developed in Python for accelerated image analysis of optical microscopy images. Batch processing can be used to quickly analyze large image sets with respect to pore size distribution, defect type, contribution of defect type to total porosity, and shape accuracy of printed samples. The open-source software is independent of the microscope used and is freely available for use. This framework allows us to perform such an analysis on a circular area with a diameter of 5 mm within 10 s, allowing detailed process maps to be obtained for new materials within minutes after preparation.

Measurement of Liver Iron Concentration in a Population of Non-Transfusion Dependent Thalassemia Patients Using a Trained Artificial Neural Network to Analyse Magnetic Resonance Images

Background: Measurements of liver iron concentration (LIC) by magnetic resonance imaging (MRI) have become established and validated in several research intensive centers. However, challenges remain with translating these measurement methodologies for routine clinical use. Current methods involve many steps presenting multiple opportunities for human error and can lead to systematic severe measurement error. Artificial neural networks (ANNs) can be trained to replace human input into image processing and may have a role in providing a simple, robust, and low cost solution for LIC measurement.Aims: The aim of the study wasto determine the limits of agreement between measurements of LIC by a reference standard spin-density-projection-assisted (SDPA) R2-MRI method (FerriScan) and a method using trained ANNs to analyse image data. MR image datasets (400) previously acquired from Siemens scanners were pre-processed and then, together with their corresponding SDPA R2-MRI LIC values, were filtered through publically available ANNs. The ANNs had been pre-trained on large image sets (ImageNet). Relevant features were extracted and then used to train a machine learning (gradient boosting) algorithm to predict LIC values from previously unseen MR image datasets.Methods: Thalassemia patients (N=100) referred by the National Institute of Haematology and Blood Transfusion, Hanoi, Vietnam for routine LIC measurement by MRI were prospectively recruited with informed consent. Patients were randomised to be scanned in either a Philips Ingenia or a Siemens Avanto 1.5T scanner. Image data were acquired using the FerriScan protocol. Images were then processed (1) using a quality controlled core laboratory data analysis service using the SDPA R2-MRI method and (2) by a non-radiologically-trained technician using the trained artificial neural network to generate LIC values. The technician was required to trace the outline of the torso and then click a mouse button. The technician using the ANN was blinded from the SDPA R2-MRI results and the analysts generating the SDPA R2-MRI results were blinded from the ANN results. Reported data were analysed using the statistical methods of Bland and Altman.Results: A plot of the ANN-predicted LIC values against the SDPA R2-MRI LIC values (Figure) shows the data scattered about the line of equivalence indicating that the ANN method is generating results that correlate with the validated SDPA R2-MRI reference standard. The geometric mean ratio of ANN LIC to SDPA R2-MRI LIC was 0.93 (95% CI 0.91 -0.96) indicating a small but statistically significant systematic underestimation of LIC by the ANN. The geometric mean ratios of the two LIC measurements were significantly different for the two scanners (0.98 for Philips and 0.90 for Siemens, p = 0.006) indicating that the bias of the ANN method against the reference standard is not universal but is dependent on scanner. Bland Altman analysis indicates that 95% of pairs of measurements are predicted to have ratios between 1.23 and 0.69 indicating a modest random variability between the ANN method and the reference standard that is better than that between biopsy and the reference standard. The performance of the ANN method in predicting SDPA R2-MRI LIC values above the clinically relevant thresholds of 7 and 15 mg Fe/g dw is characterized in Table 1 showing positive predictive values (PPVs) and negative predictive values (NPVs) together with their 95% CIs.Conclusion: The data indicate that the ANN method of measurement of LIC has potential to be used as a fast and robust method of analysing MR image data to generate LIC values. Further testing is required on a wider range of scanner types and with more patients with lower LICs. The performance of the ANN will also likely be improved through training on a larger number of datasets. [Display omitted] DisclosuresSt Pierre:Resonance Health Ltd: Consultancy, Equity Ownership, Speakers Bureau. Cooper:Resonance Health Ltd: Consultancy. Boulos:Resonance Health Ltd: Employment. House:Resonance Health Ltd: Employment. Pang:Resonance Health Ltd: Employment. Bangma:Resonance Health Ltd: Employment. Taher:Novartis Pharmaceuticals: Honoraria, Research Funding; Celgene: Research Funding.

Method G: Uncertainty Quantification for Distributed Data Problems Using Generalized Fiducial Inference

It is not unusual for a data analyst to encounter datasets distributed across several computers. This can happen for reasons such as privacy concerns, efficiency of likelihood evaluations, or just the sheer size of the whole dataset. This presents new challenges to statisticians as even computing simple summary statistics such as the median becomes computationally challenging. Furthermore, if other advanced statistical methods are desired, then novel computational strategies are needed. In this article, we propose a new approach for distributed analysis of massive data that is suitable for generalized fiducial inference and is based on a careful implementation of a “divide-and-conquer” strategy combined with importance sampling. The proposed approach requires only small amount of communication between nodes, and is shown to be asymptotically equivalent to using the whole dataset. Unlike most existing methods, the proposed approach produces uncertainty measures (such as confidence intervals) in addition to point estimates for parameters of interest. The proposed approach is also applied to the analysis of a large set of solar images. Supplementary materials for this article are available online.

Identification and Distribution of the Petrophila fulicalis Species Group (Crambidae): Taking Advantage of Citizen Science Data

Six very similar species of Petrophila Guilding are reviewed and distinguished: P. fulicalis (Clements 1860), P. confusalis (Walker, [1866]), P. canadensis (Munroe, 1972), P. santafealis (Heppner, 1976), P. hodgesi (Munroe, 1972), and P. heppneri (Blanchard & Knudson, 1983). Large digital image sets derived from citizen science databases are used to supplement traditional data sources to elucidate suites of useful diagnostic characters. These expanded data sets allow for examination of phenotypic variation across ranges and for enhanced understanding of distributional patterns. The advantages and disadvantages of such data sets are discussed and existing data gaps are identified.

Unsupervised moiré pattern removal for recaptured screen images

Moiré patterns, mainly due to the aliasing between the grids of display devices and camera sensors, often occur in recaptured screen images (videos). The spatial variety in densities and scales makes it difficult to remove moiré patterns with hand-crafted priors. Prevalent supervised-learning based methods require huge amount of paired images, which are difficult to capture and align. This paper proposes an unsupervised Generative Adversarial Network for Moiré Removal (MR-GAN), which is the first attempt at unsupervised learning based moiré removal. The recaptured images are corrected against vignetting artifacts to make MR-GAN focus on correcting brightness and contrast distortions while removing moiré patterns. We propose two complementary discriminator groups to effectively distinguish moiré patterns and image features at both large scales and small scales. A group of self-supervised loss functions, including cycle consistent loss, identity loss, cosine similarity loss, and content leakage loss, are designed to train effective generators. Experimental results demonstrate that our method outperforms state-of-the-art demoiréing methods on a large set of test images. Our code and dataset are released in the link:https://github.com/JerryLeolfl/pytorch-MRGAN-master.

Estimating Player Positions from Padel High-Angle Videos: Accuracy Comparison of Recent Computer Vision Methods.

The estimation of player positions is key for performance analysis in sport. In this paper, we focus on image-based, single-angle, player position estimation in padel. Unlike tennis, the primary camera view in professional padel videos follows a de facto standard, consisting of a high-angle shot at about 7.6 m above the court floor. This camera angle reduces the occlusion impact of the mesh that stands over the glass walls, and offers a convenient view for judging the depth of the ball and the player positions and poses. We evaluate and compare the accuracy of state-of-the-art computer vision methods on a large set of images from both amateur videos and publicly available videos from the major international padel circuit. The methods we analyze include object detection, image segmentation and pose estimation techniques, all of them based on deep convolutional neural networks. We report accuracy and average precision with respect to manually-annotated video frames. The best results are obtained by top-down pose estimation methods, which offer a detection rate of 99.8% and a RMSE below 5 and 12 cm for horizontal/vertical court-space coordinates (deviations from predicted and ground-truth player positions). These results demonstrate the suitability of pose estimation methods based on deep convolutional neural networks for estimating player positions from single-angle padel videos. Immediate applications of this work include the player and team analysis of the large collection of publicly available videos from international circuits, as well as an inexpensive method to get player positional data in amateur padel clubs.

Diagnostic accuracy of current machine learning classifiers for age-related macular degeneration: a systematic review and meta-analysis.

The objective of this study was to systematically review and meta-analyze the diagnostic accuracy of current machine learning classifiers for age-related macular degeneration (AMD). Artificial intelligence diagnostic algorithms can automatically detect and diagnose AMD through training data from large sets of fundus or OCT images. The use of AI algorithms is a powerful tool, and it is a method of obtaining a cost-effective, simple, and fast diagnosis of AMD. MEDLINE, EMBASE, CINAHL, and ProQuest Dissertations and Theses were searched systematically and thoroughly. Conferences held through Association for Research in Vision and Ophthalmology, American Academy of Ophthalmology, and Canadian Society of Ophthalmology were searched. Studies were screened using Covidence software and data on sensitivity, specificity and area under curve were extracted from the included studies. STATA 15.0 was used to conduct the meta-analysis. Our search strategy identified 307 records from online databases and 174 records from gray literature. Total of 13 records, 64,798 subjects (and 612,429 images), were used for the quantitative analysis. The pooled estimate for sensitivity was 0.918 [95% CI: 0.678, 0.98] and specificity was 0.888 [95% CI: 0.578, 0.98] for AMD screening using machine learning classifiers. The relative odds of a positive screen test in AMD cases were 89.74 [95% CI: 3.05-2641.59] times more likely than a negative screen test in non-AMD cases. The positive likelihood ratio was 8.22 [95% CI: 1.52-44.48] and the negative likelihood ratio was 0.09 [95% CI: 0.02-0.52]. The included studies show promising results for the diagnostic accuracy of the machine learning classifiers for AMD and its implementation in clinical settings.

Explaining face representation in the primate brain using different computational models

Understanding how the brain represents the identity of complex objects is a central challenge of visual neuroscience. The principles governing object processing have been extensively studied in the macaque face patch system, a sub-network of inferotemporal (IT) cortex specialized for face processing. A previous study reported that single face patch neurons encode axes of a generative model called the "active appearance" model, which transforms 50D feature vectors separately representing facial shape and facial texture into facial images. However, a systematic investigation comparing this model to other computational models, especially convolutional neural network models that have shown success in explaining neural responses in the ventral visual stream, has been lacking. Here, we recorded responses of cells in the most anterior face patch anterior medial (AM) to a large set of real face images and compared a large number of models for explaining neural responses. We found that the active appearance model better explained responses than any other model except CORnet-Z, a feedforward deep neural network trained on general object classification to classify non-face images, whose performance it tied on some face image sets and exceeded on others. Surprisingly, deep neural networks trained specifically on facial identification did not explain neural responses well. A major reason is that units in the network, unlike neurons, are less modulated by face-related factors unrelated to facial identification, such as illumination.

LANDSLIDE SUSCEPTIBILITY ASSESSMENT IN JIUZHAIGOU SCENIC AREA WITH GIS BASED ON CERTAINTY FACTOR AND LOGISTIC REGRESSION MODEL

Thousands of coseismic landslides were induced by the Jiuzhaigou MS7.0 earthquake on August 8th 2017, Sichuan, China. Thus, a tremendous volume of loose debris are extremely prone to convert into landslides and debris flow under rainfall. Therefore, it is necessary to conduct landslide susceptibility assessment in Jiuzhaigou Scenic area. Based on a large set of pre-and post-earthquake high-resolution satellite images and field investigation, a polygon-based interpretation of the coseismic landslides is carried out. In total, 1047 landslides are identified, covering an area of 3.88 km2. The relationship between the spatial distribution of landslides and its influencing factors is analyzed. Nine factors including tectonic, topographic, geological and other factors are selected to evaluate the landslide susceptibility by certainty factor model and logistic regression model, as well as coupled model. The result shows that slope, aspect, elevation and lithology are the main factors controlling the spatial distribution of landslides. The susceptibility of landslide is divided into low(60.72%), moderate(24.18%), high(9.89%) and extremely high areas(5.21%). The high and extremely high areas covering an area of 99 km2 including Tiger Lake and Panda Lake are distributed along the valleys. The assessment result of the coupled model is more reasonable than that of the single model and can be used as scientific reference for prevention of landslide hazard and segmented opening time of scenic area.

Automated Quantitative Analysis of ex vivo Blood-Brain Barrier Permeability Using Intellesis Machine-Learning.

BackgroundAn increase in blood brain barrier permeability commonly precedes neuro-inflammation and cognitive impairment in models of dementia. Common methods to estimate capillary permeability have potential confounders, or require laborious and subjective semi-manual analysis.New methodHere we used snap frozen mouse and rat brain sections that were double-immunofluorescent labeled for immunoglobulin G (IgG; plasma protein) and laminin-α4 (capillary basement membrane). A Machine Learning Image Analysis program (Zeiss ZEN Intellisis) was trained to recognize and segment laminin-α4 to equivocally identify blood vessels in large sets of images. An IgG subclass based on a threshold intensity was segmented and quantitated only in extravascular regions. The residual parenchymal IgG fluorescence is indicative of blood-to-brain extravasation of IgG and was accurately quantitated.ResultsAutomated machine-learning and threshold based segmentation of only parenchymal IgG extravasation accentuates otherwise indistinct capillary permeability, particularly frequent in minor BBB leakage. Comparison with Existing Methods: Large datasets can be processed and analyzed quickly and robustly to provide an overview of vascular permeability throughout the brain. All human bias or ambiguity involved in classifying and measuring leakage is removed.ConclusionHere we describe a fast and precise method of visualizing and quantitating BBB permeability in mouse and rat brain tissue, while avoiding the confounding influence of unphysiological conditions such as perfusion and eliminating any human related bias from analysis.

Similarity learning to enable building searches in post‐event image data

AbstractReconnaissance teams collect perishable data after each disaster to learn about building performance. However, often these large image sets are not adequately curated, nor do they have sufficient metadata (e.g., GPS), hindering any chance to identify images from the same building when collected by different reconnaissance teams. In this study, Siamese convolutional neural networks (S‐CNN) are implemented and repurposed to establish a building search capability suitable for post‐disaster imagery. This method can automatically rank and retrieve corresponding building images in response to a single query using an image. In the demonstration, we utilize real‐world images collected from 174 reinforced‐concrete buildings affected by the 2016 Southern Taiwan and the 2017 Pohang (South Korea) earthquake events. A quantitative performance evaluation is conducted by examining two metrics introduced for this application: Similarity Score (SS) and Similarity Rank (SR).

A Progressive and Cross-Domain Deep Transfer Learning Framework for Wrist Fracture Detection

Abstract There has been an amplified focus on and benefit from the adoption of artificial intelligence (AI) in medical imaging applications. However, deep learning approaches involve training with massive amounts of annotated data in order to guarantee generalization and achieve high accuracies. Gathering and annotating large sets of training images require expertise which is both expensive and time-consuming, especially in the medical field. Furthermore, in health care systems where mistakes can have catastrophic consequences, there is a general mistrust in the black-box aspect of AI models. In this work, we focus on improving the performance of medical imaging applications when limited data is available while focusing on the interpretability aspect of the proposed AI model. This is achieved by employing a novel transfer learning framework, progressive transfer learning, an automated annotation technique and a correlation analysis experiment on the learned representations. Progressive transfer learning helps jump-start the training of deep neural networks while improving the performance by gradually transferring knowledge from two source tasks into the target task. It is empirically tested on the wrist fracture detection application by first training a general radiology network RadiNet and using its weights to initialize RadiNetwrist , that is trained on wrist images to detect fractures. Experiments show that RadiNetwrist achieves an accuracy of 87% and an AUC ROC of 94% as opposed to 83% and 92% when it is pre-trained on the ImageNet dataset. This improvement in performance is investigated within an explainable AI framework. More concretely, the learned deep representations of RadiNetwrist are compared to those learned by the baseline model by conducting a correlation analysis experiment. The results show that, when transfer learning is gradually applied, some features are learned earlier in the network. Moreover, the deep layers in the progressive transfer learning framework are shown to encode features that are not encountered when traditional transfer learning techniques are applied. In addition to the empirical results, a clinical study is conducted and the performance of RadiNetwrist is compared to that of an expert radiologist. We found that RadiNetwrist exhibited similar performance to that of radiologists with more than 20 years of experience. This motivates follow-up research to train on more data to feasibly surpass radiologists’ performance, and investigate the interpretability of AI models in the healthcare domain where the decision-making process needs to be credible and transparent.