Deep Neural Networks For Detection Research Articles

Developing the Framework Using Deep Neural Network for Detection of Spam and Fake Spam Messages in Twitter

Social media plays vital role among the user communities for social gathering, entertainment, communication, sharing knowledge so on. Twitter is one such network to connect millions of users to share information. Nowadays, there are humpteen numbers of users using social media for social engagements. Due to the fact that wide publicity of individuals and products get viral in social media, everyone wish to use social media as a platform to promote their product. Furthermore, large number of people relies on social media contents to take decisions. Twitter is one of the social media platforms to post the media contents by the user. Spammers are illegal users intrude the twitter account and send the duplicate messages to promote advertisement, phishing, scam and personal blogs etc. In this paper, a novel spam detection mechanism is introduced to detect the suspicious users on twitter. The system has been designed such a way that it initially set with semi-supervised at the tweet level and update into supervised level for learning the input tweets to detect the spammers. The proposed system will also identify the type of spammers and will also remove duplicate tweets. We have applied with multi-classifier algorithms like naïve Bayesian, K-Nearest neighbor and Random forest into twitter data set and the performance is compared. The experimental result shows very promising results.

Exploring Semantic Redundancy using Backdoor Triggers: A Complementary Insight into the Challenges facing DNN-based Software Vulnerability Detection

To detect software vulnerabilities with better performance, deep neural networks (DNNs) have received extensive attention recently. However, these vulnerability detection DNN models trained with code representations are vulnerable to specific perturbations on code representations. This motivates us to rethink the bane of software vulnerability detection and find function-agnostic features during code representation which we name as semantic redundant features. This paper first identifies a tight correlation between function-agnostic triggers and semantic redundant feature space (where the redundant features reside) in these DNN models. For correlation identification, we propose a novel Backdoor-based Semantic Redundancy Exploration (BSemRE) framework. In BSemRE, the sensitivity of the trained models to function-agnostic triggers is observed to verify the existence of semantic redundancy in various code representations. Specifically, acting as the typical manifestations of semantic redundancy, naming conventions, ternary operators and identically-true conditions are exploited to generate function-agnostic triggers. Extensive comparative experiments on 1613823 samples of 8 representative vulnerability datasets and state-of-the-art code representation techniques and vulnerability detection models demonstrate that the existence of semantic redundancy determines the upper trustworthiness limit of DNN-based software vulnerability detection. To the best of our knowledge, this is the first work exploring the bane of software vulnerability detection using backdoor triggers.

PArtNNer: Platform-Agnostic Adaptive Edge-Cloud DNN Partitioning for Minimizing End-to-End Latency

The last decade has seen the emergence of Deep Neural Networks (DNNs) as the de facto algorithm for various computer vision applications. In intelligent edge devices, sensor data streams acquired by the device are processed by a DNN application running on either the edge device itself or in the cloud. However, “edge-only” and “cloud-only” execution of State-of-the-Art DNNs may not meet an application’s latency requirements due to the limited compute, memory, and energy resources in edge devices, dynamically varying bandwidth of edge-cloud connectivity networks, and temporal variations in the computational load of cloud servers. This work investigates distributed (partitioned) inference across edge devices (mobile/end device) and cloud servers to minimize end-to-end DNN inference latency. We study the impact of temporally varying operating conditions and the underlying compute and communication architecture on the decision of whether to run the inference solely on the edge, entirely in the cloud, or by partitioning the DNN model execution among the two. Leveraging the insights gained from this study and the wide variation in the capabilities of various edge platforms that run DNN inference, we propose PArtNNer , a platform-agnostic adaptive DNN partitioning algorithm that finds the optimal partitioning point in DNNs to minimize inference latency. PArtNNer can adapt to dynamic variations in communication bandwidth and cloud server load without requiring pre-characterization of underlying platforms. Experimental results for six image classification and object detection DNNs on a set of five commercial off-the-shelf compute platforms and three communication standards indicate that PArtNNer results in 10.2× and 3.2× (on average) and up to 21.1× and 6.7× improvements in end-to-end inference latency compared to execution of the DNN entirely on the edge device or entirely on a cloud server, respectively. Compared to pre-characterization-based partitioning approaches, PArtNNer converges to the optimal partitioning point 17.6× faster.

High-Throughput Spike Detection in Greenhouse Cultivated Grain Crops with Attention Mechanisms-Based Deep Learning Models.

Detection of spikes is the first important step toward image-based quantitative assessment of crop yield. However, spikes of grain plants occupy only a tiny fraction of the image area and often emerge in the middle of the mass of plant leaves that exhibit similar colors to spike regions. Consequently, accurate detection of grain spikes renders, in general, a non-trivial task even for advanced, state-of-the-art deep neural networks (DNNs). To improve pattern detection in spikes, we propose architectural changes to Faster-RCNN (FRCNN) by reducing feature extraction layers and introducing a global attention module. The performance of our extended FRCNN-A vs. conventional FRCNN was compared on images of different European wheat cultivars, including "difficult" bushy phenotypes from 2 different phenotyping facilities and optical setups. Our experimental results show that introduced architectural adaptations in FRCNN-A helped to improve spike detection accuracy in inner regions. The mean average precision (mAP) of FRCNN and FRCNN-A on inner spikes is 76.0% and 81.0%, respectively, while on the state-of-the-art detection DNNs, Swin Transformer mAP is 83.0%. As a lightweight network, FRCNN-A is faster than FRCNN and Swin Transformer on both baseline and augmented training datasets. On the FastGAN augmented dataset, FRCNN achieved a mAP of 84.24%, FRCNN-A attained a mAP of 85.0%, and the Swin Transformer achieved a mAP of 89.45%. The increase in mAP of DNNs on the augmented datasets is proportional to the amount of the IPK original and augmented images. Overall, this study indicates a superior performance of attention mechanisms-based deep learning models in detecting small and subtle features of grain spikes.

Hierarchical waste detection with weakly supervised segmentation in images from recycling plants

Reducing environmental pollution with household waste and emissions from the computing clusters is an urgent technological problem. In our work, we explore both of these aspects: the deep learning application to improve the efficiency of waste recognition on recycling plant’s conveyor, as well as carbon dioxide emission from the computing devices used in this process. To conduct research, we developed an unique open WaRP dataset that demonstrates the best diversity among similar industrial datasets and contains more than 10,000 images with 28 different types of recyclable goods (bottles, glasses, card boards, cans, detergents, and canisters). Objects can overlap, be in poor lighting conditions, or significantly distorted. On the WaRP dataset, we study training and evaluation of cutting-edge deep neural networks for detection, classification and segmentation tasks. Additionally, we developed a hierarchical neural network approach called H-YC with weakly supervised waste segmentation. It provided a notable increase in the detection quality and made it possible to segment images, learning only having class labels, not their masks. Both the suggested hierarchical approach and the WaRP dataset have shown great industrial application potential.

Abstract 13945: Evaluation of a Deep Neural Network for Detection of D-Transposition of the Great Arteries on Fetal Echocardiograms

Introduction: While prenatal detection of congenital heart disease (CHD) has improved over the years, a significant proportion of CHD remains unrecognized on fetal ultrasound. Prenatal identification of d-transposition of the great arteries (d-TGA) in particular would significantly alter delivery planning and postnatal course. Research Questions: We investigated whether a deep neural network (DNN) can accurately detect d-TGA on fetal echocardiograms. Aim: To evaluate a DNN for detection of d-TGA on second (2T) or third (3T) trimester fetal echocardiograms. Methods: Patients with a singleton pregnancy with a 2T or 3T fetal echocardiogram performed between Jan 1, 2022 and May 1, 2023 at one center were retrospectively evaluated. Cases of d-TGA and negative cases referred for family history of CHD or increased nuchal translucency were included. Cases were reviewed by 3 pediatric cardiologists to confirm that the recorded ultrasound clips were sufficient to diagnose the presence or absence of d-TGA, and were excluded otherwise. The DNN was trained on patients not included in the evaluation to detect the absence of the great artery cross-over pattern, a distinct finding in d-TGA. The DNN outputs the presence or absence of this finding using all recorded video clips of an examination. Results: We included 18 positive and 99 negative cases. The DNN detected suspicion for d-TGA with an AUC of 91.4%, sensitivity of 83.3% (60.8-94.2, 95%CI) and specificity of 92.9% (86.1-96.5, 95%CI). Conclusions: D-TGA can be accurately detected on fetal echocardiograms using a DNN. These promising results lay the foundation for reliable, computer-aided detection of CHD on screening fetal ultrasounds.

Energy-Efficient Approximate Edge Inference Systems

The rapid proliferation of the Internet of Things and the dramatic resurgence of artificial intelligence based application workloads have led to immense interest in performing inference on energy-constrained edge devices. Approximate computing (a design paradigm that trades off a small degradation in application quality for disproportionate energy savings) is a promising technique to enable energy-efficient inference at the edge. This article introduces the concept of an approximate edge inference system ( AxIS ) and proposes a systematic methodology to perform joint approximations between different subsystems in a deep neural network (DNN)-based edge inference system, leading to significant energy benefits compared to approximating individual subsystems in isolation. We use a smart camera system that executes various DNN-based image classification and object detection applications to illustrate how the sensor, memory, compute, and communication subsystems can all be approximated synergistically. We demonstrate our proposed methodology using two variants of a smart camera system: (a) Cam Edge , where the DNN is executed locally on the edge device, and (b) Cam Cloud , where the edge device sends the captured image to a remote cloud server that executes the DNN. We have prototyped such an approximate inference system using an Intel Stratix IV GX-based Terasic TR4-230 FPGA development board. Experimental results obtained using six large DNNs and four compact DNNs running image classification applications demonstrate significant energy savings (≈ 1.6× -4.7× for large DNNs and ≈ 1.5× -3.6× for small DNNs), for minimal (<1%) loss in application-level quality. Furthermore, results using four object detection DNNs exhibit energy savings of ≈ 1.5× -5.2× for similar quality loss. Compared to approximating a single subsystem in isolation, AxIS achieves 1.05× -3.25× gains in energy savings for image classification and 1.35× -4.2× gains for object detection on average, for minimal (<1%) application-level quality loss.

Signals of global symmetry are important for abnormality detection in mammograms.

Expert radiologists can detect the "gist of abnormal" in bilateral mammograms even three years prior to onset of cancer. However, their performance decreases if both breasts are not from the same woman, suggesting the ability to detect the abnormality is partly dependent on a global signal present across the two breasts. We aim to detect this implicitly perceived "symmetry" signal by examining its effect on a pre-trained mammography model. A deep neural network (DNN) with four mammogram view inputs was developed to predict whether the mammograms come from one woman, or two different women as the first step in investigating the symmetry signal. Mammograms were balanced by size, age, density, and machine type. We then evaluated a cancer detection DNN's performance on mammograms from the same and different women. Finally, we used textural analysis methods to further explain the symmetry signal. The developed DNN can detect whether a set of mammograms come from the same or different woman with a base accuracy of 61%. Indeed, a DNN shown mammograms swapped either contralateral or abnormal with a normal mammogram from another woman, resulted in performance decreases. Findings indicate that abnormalities induce a disruption in global mammogram structure resulting in the break in the critical symmetry signal. The global symmetry signal is a textural signal embedded in the parenchyma of bilateral mammograms, which can be extracted. The presence of abnormalities alters textural similarities between the left and right breasts and contributes to the "medical gist signal."

Establishment and evaluation of conditional GAN-based image dataset for semantic segmentation of structural cracks

Crack detection is vital to maintain the structural safety of in-service bridges, which is an increasing demand in the industrial community. The deep learning-based crack detection is an emerging method that provides a novel way to deal with this problem. Although, the deep neural networks (DNN) can learn to detect cracks themselves, they require large numbers of crack images to learn the features of cracks in the real world. Besides, except for the crack forms, there always exist all kinds of noise motifs that will disturb the correct detection of crack regions. The lack of crack sample images has been an obstacle for the improvement of deep learning-based crack detection method. This paper proposes a generative adversarial network (GAN) based method to establish a synthesized crack image dataset with pixel-wise annotations, which provides a novel way aside from the traditional data augmentation method. The Deep Convolutional GAN (DCGAN) model was adopted for the generation of synthesized crack annotations while the Pixel2Pixel model was used to generate the corresponding synthesized crack images. The generated crack annotations and crack images during the training epochs were demonstrated to show how the GANs learn to generate the synthesized images. Moreover, comparative study was conducted to validate the performance of the synthesized crack image dataset for training the crack detection DNN. The results showed the DNN trained by the synthesized images can achieve 74.34% of the MeanIoU that was reached by the same DNN model trained with real images. As for the way of using the synthesized and real crack images for training a crack detection DNN, the way that use the synthesized crack images for pre-training and then use the real images for fine-tuning is better than the way that directly mix the synthesized and real crack images for training. This work provides reference for the GAN-based establishment of crack image dataset and the evaluation of the image quality for training crack detection DNNs.

A transformer‐based deep neural network for detection and classification of lung cancer via PET/CT images

AbstractLung cancer is the leading cause of death for men and women worldwide and the second most frequent cancer. Therefore, early detection of the disease increases the cure rate. This paper presents a new approach to evaluate the ability of positron emission tomography/computed tomography (PET/CT) images to classify and detect lung cancer using deep learning techniques. Our approach aims to fully automate lung cancer's anatomical localization from PET/CT images. It also searches to classify the tumor, which is essential as it makes it possible to determine the disease's speed of progression and the best treatments to adopt. We have built, in this work, an approach based on transformers by implementing the DETR model as a tool to detect the tumor and assist physicians in staging patients with lung cancer. The TNM staging system and histologic subtype classification were both taken as a standard for classification. Experimental results demonstrated that our approach achieves sound results on tumor localization, T staging, and histology classification. Our proposed approach detects tumors with an intersection over union (IOU) of 0.8 when tested on the Lung‐PET‐CT‐Dx dataset. It also has yielded better accuracy than state‐of‐the‐art T‐staging and histologic classification methods. It classified T‐stage and histologic subtypes with an accuracy of 0.97 and 0.94, respectively.

Comparative Analysis of Artificial Neural Networks and Deep Neural Networks for Detection of Dementia

Dementia is a neurocognitive brain disease that emerged as a worldwide health challenge. Machine learning and deep learning have been effectively applied for the detection of dementia using magnetic resonance imaging. In this work, the performance of both machine learning and deep learning frameworks along with artificial neural networks are assessed for detecting dementia and normal subjects using MRI images. The first-order and second-order hand-crafted features are used as input for machine learning and artificial neural networks. And automatic feature extraction is used in the last framework with the pre-trained networks. The outcomes show that the framework using the deep neural networks performs better contrasted with the first two methodologies used in terms of various performance measures.

A novel deep neural network for detection of Atrial Fibrillation using ECG signals

In healthcare practice, one of the most predominantly occurring dysrhythmia is atrial fibrillation (AF). The anomalous heart rhythm and the deficiency of an evident P-wave signal are the consequences of AF, several cerebral apoplexies, thrombus, blood coagulation, cognitive impairments, and strokes. It is arduous to ascertain the symptoms of AF and clinically silent that might cause death. There are certain liabilities for diagnosis of AF in manual Electrocardiogram (ECG) since it demands high expertise; it is a time demanding and tedious process which is also accompanied by variations between intra- and inter-observer. Hence, to combat with this issue, a novel AF detection models has been proposed a conglomerate parallel structure of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) framework which can deepen the understanding of features and classification. To test and legitimize the system, we utilize the data of the MIT-BIH Atrial Fibrillation Database. Parallel models for control subjects have been designed specifically to validate performance in terms of classification for more voracious categorization of 3 classes namely: Non-Atrial Fibrillation (N-AF), Atrial Fibrillation (AF) and Normal Sinus Rhythm (NSR). The model obtained an Accuracy, Sensitivity and Specificity of 99.6%, 98.64% and 99.01% respectively.

A lightweight deep neural network for detection of mental states from physiological signals

A lightweight deep neural network for detection of mental states from physiological signals

COVID-Net CT-2: Enhanced Deep Neural Networks for Detection of COVID-19 From Chest CT Images Through Bigger, More Diverse Learning.

The COVID-19 pandemic continues to rage on, with multiple waves causing substantial harm to health and economies around the world. Motivated by the use of computed tomography (CT) imaging at clinical institutes around the world as an effective complementary screening method to RT-PCR testing, we introduced COVID-Net CT, a deep neural network tailored for detection of COVID-19 cases from chest CT images, along with a large curated benchmark dataset comprising 1,489 patient cases as part of the open-source COVID-Net initiative. However, one potential limiting factor is restricted data quantity and diversity given the single nation patient cohort used in the study. To address this limitation, in this study we introduce enhanced deep neural networks for COVID-19 detection from chest CT images which are trained using a large, diverse, multinational patient cohort. We accomplish this through the introduction of two new CT benchmark datasets, the largest of which comprises a multinational cohort of 4,501 patients from at least 16 countries. To the best of our knowledge, this represents the largest, most diverse multinational cohort for COVID-19 CT images in open-access form. Additionally, we introduce a novel lightweight neural network architecture called COVID-Net CT S, which is significantly smaller and faster than the previously introduced COVID-Net CT architecture. We leverage explainability to investigate the decision-making behavior of the trained models and ensure that decisions are based on relevant indicators, with the results for select cases reviewed and reported on by two board-certified radiologists with over 10 and 30 years of experience, respectively. The best-performing deep neural network in this study achieved accuracy, COVID-19 sensitivity, positive predictive value, specificity, and negative predictive value of 99.0%/99.1%/98.0%/99.4%/99.7%, respectively. Moreover, explainability-driven performance validation shows consistency with radiologist interpretation by leveraging correct, clinically relevant critical factors. The results are promising and suggest the strong potential of deep neural networks as an effective tool for computer-aided COVID-19 assessment. While not a production-ready solution, we hope the open-source, open-access release of COVID-Net CT-2 and the associated benchmark datasets will continue to enable researchers, clinicians, and citizen data scientists alike to build upon them.

Adaptive Anomaly Detection for Internet of Things in Hierarchical Edge Computing: A Contextual-Bandit Approach

The advances in deep neural networks (DNN) have significantly enhanced real-time detection of anomalous data in IoT applications. However, the complexity-accuracy-delay dilemma persists: Complex DNN models offer higher accuracy, but typical IoT devices can barely afford the computation load, and the remedy of offloading the load to the cloud incurs long delay. In this article, we address this challenge by proposing an adaptive anomaly detection scheme with hierarchical edge computing (HEC). Specifically, we first construct multiple anomaly detection DNN models with increasing complexity and associate each of them to a corresponding HEC layer. Then, we design an adaptive model selection scheme that is formulated as a contextual-bandit problem and solved by using a reinforcement learning policy network . We also incorporate a parallelism policy training method to accelerate the training process by taking advantage of distributed models. We build an HEC testbed using real IoT devices and implement and evaluate our contextual-bandit approach with both univariate and multivariate IoT datasets. In comparison with both baseline and state-of-the-art schemes, our adaptive approach strikes the best accuracy-delay tradeoff on the univariate dataset and achieves the best accuracy and F1-score on the multivariate dataset with only negligibly longer delay than the best (but inflexible) scheme.

Deep Neural Networks for Detection and Location of Microseismic Events and Velocity Model Inversion from Microseismic Data Acquired by Distributed Acoustic Sensing Array.

Fiber-optic cables have recently gained popularity for use as Distributed Acoustic Sensing (DAS) arrays for borehole microseismic monitoring due to their physical robustness as well as high spatial and temporal resolutions. As a result, the sensors record large amounts of data, making it very difficult to process in real-/semi-real-time using the conventional processing routines. We present a novel approach, based on deep learning, for handling the large amounts of DAS data in real-/semi-real-time. The proposed neural network was trained on synthetic microseismic data contaminated with real-ambient noise from field data and was validated using field DAS microseismic data obtained from a hydraulic fracturing operation. The results indicate that the trained network is capable of detecting and locating microseismic events from DAS data and simultaneously update the velocity model to a high degree of precision. The mean absolute errors in the event locations and the velocity model parameters are 2.04, 0.72, 2.76, 4.19 and 0.97 percent for distance (x), depth (z), P-wave velocity, S-wave velocity and density, respectively. In addition to automation and computational efficiency, deep learning reduces human expert data handling during processing, thus preserving data integrity leading to more accurate and reproducible results.

Comparative Analysis of Deep Neural Networks for the Detection and Decoding of Data Matrix Landmarks in Cluttered Indoor Environments

Data Matrix patterns imprinted as passive visual landmarks have shown to be a valid solution for the self-localization of Automated Guided Vehicles (AGVs) in shop floors. However, existing Data Matrix decoding applications take a long time to detect and segment the markers in the input image. Therefore, this paper proposes a pipeline where the detector is based on a real-time Deep Learning network and the decoder is a conventional method, i.e. the implementation in libdmtx. To do so, several types of Deep Neural Networks (DNNs) for object detection were studied, trained, compared, and assessed. The architectures range from region proposals (Faster R-CNN) to single-shot methods (SSD and YOLO). This study focused on performance and processing time to select the best Deep Learning (DL) model to carry out the detection of the visual markers. Additionally, a specific data set was created to evaluate those networks. This test set includes demanding situations, such as high illumination gradients in the same scene and Data Matrix markers positioned in skewed planes. The proposed approach outperformed the best known and most used Data Matrix decoder available in libraries like libdmtx.

A New Smartphone-Based Optic Nerve Head Biometric for Verification and Change Detection.

PurposeLens adapted smartphones are being used regularly instead of ophthalmoscopes. The most common causes of preventable blindness in the world, which are glaucoma and diabetic retinopathy, can develop asymptomatic changes to the optic nerve head (ONH) especially in the developing world where there is a dire shortage of ophthalmologists but ubiquitous mobile phones. We developed a proof-of-concept ONH biometric (application [APP]) to use as a routine biometric on a mobile phone. The unique blood vessel pattern is verified if it maps on to a previously enrolled image.MethodsThe iKey APP platform comprises three deep neural networks (DNNs) developed from anonymous ONH images: the graticule blood vessel (GBV) and the blood vessel specific feature (BVSF) DNNs were trained on unique blood vessel vectors. A non-feature specific (NFS) baseline ResNet50 DNN was trained for comparison.ResultsVerification reached an accuracy of 97.06% with BVSF, 87.24% with GBV and 79.8% using NFS.ConclusionsA new ONH biometric was developed with a hybrid platform of ONH algorithms for use as a verification biometric on a smartphone. Failure to verify will alert the user to possible changes to the image, so that silent changes may be observed before sight threatening disease progresses. The APP retains a history of all ONH images. Future longitudinal analysis will explore the impact of ONH changes to the iKey biometric platform.Translational RelevancePhones with iKey will host ONH images for biometric protection of both health and financial data. The ONH may be used for automatic screening by new disease detection DNNs.

Extraction of linear structures from digital terrain models using deep learning

Abstract. This paper explores the role deep convolutional neural networks play in automated extraction of linear structures using semantic segmentation techniques in Digital Terrain Models (DTMs). DTM is a regularly gridded raster created from laser scanning point clouds and represents elevations of the bare earth surface with respect to a reference. Recent advances in Deep Learning (DL) have made it possible to explore the use of semantic segmentation for detection of terrain structures in DTMs. This research examines two novel and practical deep convolutional neural network architectures i.e. an encoder-decoder network named as SegNet and the recent state-of-the-art high-resolution network (HRNet). This paper initially focuses on the pixel-wise binary classification in order to validate the applicability of the proposed approaches. The networks are trained to distinguish between points belonging to linear structures and those belonging to background. In the second step, multi-class segmentation is carried out on the same DTM dataset. The model is trained to not only detect a linear feature, but also to categorize it as one of the classes: hollow ways, roads, forest paths, historical paths, and streams. Results of the experiment in addition to the quantitative and qualitative analysis show the applicability of deep neural networks for detection of terrain structures in DTMs. From the deep learning models utilized, HRNet gives better results.

Ensembled Deep Neural Network for Intracranial Hemorrhage Detection and Subtype Classification on Noncontrast CT Images

Rapid and accurate diagnosis of intracranial hemorrhage is clinically significant to ensure timely treatment. In this study, we developed an ensembled deep neural network for the detection and subtype classification of intracranial hemorrhage. The model consisted of two parallel network pathways, one using three different window level/width settings to enhance the image contrast of brain, blood, and soft tissue. The other extracted spatial information of adjacent image slices to the target slice. Both pathways exploited the EfficientNet-B0 as the basic architecture and were ensembled to generate the final prediction. Class activation mapping was applied in both pathways to highlight the regions of detected hemorrhage and the associated subtypes. The model was trained and tested using Intracranial Hemorrhage Detection Challenge (IHDC) dataset launched by the Radiological Society of North America (RSNA) in 2019, which contained 674,258 head noncontrasts computer tomography images acquired from 19,530 patients. An independent dataset (CQ500) acquired from another institution was used to test the generalizability of the trained model. The overall accuracy, sensitivity, and F1 score for intracranial hemorrhage detection were 95.7%, 85.9%, and 86.7% on IHDC testing dataset and 92.4%, 92.6%, and 93.4% on external CQ500 testing dataset. The heatmaps by class activation mapping successfully demonstrated discriminative feature regions of the predicted hemorrhage locations and subtypes, providing visual guidance for radiologists to assist in rapid diagnosis of intracranial hemorrhage.