VGG16 Network Research Articles

An effective deep learning network for detecting and classifying glaucomatous eye

<span lang="EN-US">Glaucoma is a well-known complex disease of the optic nerve that gradually damages eyesight due to the increase of intraocular pressure inside the eyes. Among two types of glaucoma, open-angle glaucoma is mostly happened by high intraocular pressure and can damage the eyes temporarily or sometimes permanently, another one is angle-closure glaucoma. Therefore, being diagnosed in the early stage is necessary to safe our vision. There are several ways to detect glaucomatous eyes like tonometry, perimetry, and gonioscopy but require time and expertise. Using deep learning approaches could be a better solution. This study focused on the recognition of open-angle affected eyes from the fundus images using deep learning techniques. The study evolved by applying VGG16, VGG19, and ResNet50 deep neural network architectures for classifying glaucoma positive and negative eyes. The experiment was executed on a public dataset collected from Kaggle; however, every model performed better after augmenting the dataset, and the accuracy was between 93% and 97.56%. Among the three models, VGG19 achieved the highest accuracy at 97.56%.</span>

PithaNet: A transfer learning-based approach for traditional pitha classification

<span lang="EN-US">Pitha, pithe, or peetha are all Bangla words referring to a native and traditional food of Bangladesh as well as some areas of India, especially the parts of India where Bangla is the primary language. Numerous types of pithas exist in the culture and heritage of the Bengali and Bangladeshi people. Pithas are traditionally prepared and offered on important occasions in Bangladesh, such as welcoming a bride grooms, or bride, entertaining guests, or planning a special gathering of family, relatives, or friends. The traditional pitha celebration and pitha culture are no longer widely practiced in modern civilization. Consequently, the younger generation is unfamiliar with our traditional pitha culture. In this study, an effective pitha image classification system is introduced. convolutional neural network (CNN) pre-trained models EfficientNetB6, ResNet50, and VGG16 are used to classify the images of pitha. The dataset of traditional popular pithas is collected from different parts of Bangladesh. In this experiment, EfficientNetB6 and ResNet50 show nearly 90% accuracy. The best classification result was obtained using VGG16 with 92% accuracy. The main motive of this study is to revive the Bengali pitha tradition among young people and people worldwide, which will encourage many other researchers to pursue research in this domain.</span>

Identification of lung disease types using convolutional neural network and VGG-16 architecture

Pneumonia, tuberculosis, and Covid-19 are different lung diseases but have similar characteristics. One of the reasons for the worsening of disease in lung sufferers is a diagnosis that takes a long time. Another factor, the results of the X-ray photos look blurry and lack contracture, causing different diagnostic results of X-ray photos. This research classifies lung images into four categories: normal lungs, tuberculosis, pneumonia, and Covid-19 using the Convolutional Neural Network method and VGG-16 architecture. The results of the research with models and scenarios without pre-trained use data with a ratio of 9:1 at epoch 50, an accuracy of 94%, while the lowest results are in scenarios using data with a ratio of 8:2 at epoch 50, non-pre-trained models, accuracy by 87%.

APPLICATION OF DEEP LEARNING TECHNIQUES IN IDENTIFICATION OF THE STRUCTURE OF SELECTED ROAD MATERIALS

In research, there is a growing interest in using artificial intelligence to find solutions to difficult scientific problems. In this paper, a deep learning algorithm has been applied using images of samples of materials used for road surfaces. The photographs showed cross-sections of random samples taken with a CT scanner. Historical samples were used for the analysis, located in a database collecting information over many years. The deep learning analysis was performed using some elements of the VGG16 network architecture and implemented using the R language. The learning and training data were augmented and cross-validated. This resulted in the high level of 96.4% quality identification of the sample type and its selected structural features. The photographs in the identification set were correctly identified in terms of structure, mix type and grain size. The trained model identified samples in the domain of the dataset used for training in a very good way. As a result, in the future such a methodology may facilitate the identification of the type of mixture, its basic properties and defects.

Ancient Architectural Damage Recognition Based on AlexNet and Transfer Learning

Due to the need for a large dataset to train image recognition using convolutional neural networks, obtaining accurately categorized images of ancient architectural structure damage poses certain challenges. To improve the recognition accuracy of damage in ancient architectural structures, we have enhanced the Vgg-16 network by combining transfer learning techniques with convolutional neural networks. This design, a new image recognition method based on transfer learning, effectively achieves precise identification of structural damage in ancient architecture, showing potential for real-world applications.

Nondestructive identification of soybean protein in minced chicken meat based on hyperspectral imaging and VGG16-SVM

A rapid and nondestructive method for identification of soybean protein in minced chicken meat based on visible-near infrared hyperspectral imaging (HSI) technology and VGG16-SVM model was proposed in this study. The hyperspectral data from 400.89 nm to 1000.19 nm of samples with different soybean protein contents were collected. Spectral data was transformed into spectrograms through Continuous wavelet transform (CWT). Comparing with other methods, the optimum data preprocessing results were obtained for the subsequent research. Then, VGG16-SVM model based on VGG16 network in which fully connected layer was replaced by support vector machine (SVM) was established to identify soybean protein in minced chicken meat. The results showed that VGG16-SVM model with radial basis function (RBF) as kernel function performed best and the classification accuracy reached 98.1%. Meanwhile, the proposed model was superior to the SVM, convolution neural network (CNN) and VGG16 model. It is feasible and effective for combining HSI technology and VGG16-SVM to realize nondestructive identification of soybean protein in minced chicken meat.

XCovNet: An optimized xception convolutional neural network for classification of COVID-19 from point-of-care lung ultrasound images

Global livelihoods are impacted by the novel coronavirus (COVID-19) disease, which mostly affects the respiratory system and spreads via airborne transmission. The disease has spread to almost every nation and is still widespread worldwide. Early and reliable diagnosis is essential to prevent the development of this highly risky disease. The computer-aided diagnostic model facilitates medical practitioners in obtaining a quick and accurate diagnosis. To address these limitations, this study develops an optimized Xception convolutional neural network, called "XCovNet," for recognizing COVID-19 from point-of-care ultrasound (POCUS) images. This model employs a stack of modules, each of which has a slew of feature extractors that enable it to learn richer representations with fewer parameters. The model identifies the presence of COVID-19 by classifying POCUS images containing Coronavirus samples, viral pneumonia samples, and healthy ultrasound images. We compare and evaluate the proposed network with state-of-the-art (SOTA) deep learning models such as VGG, DenseNet, Inception-V3, ResNet, and Xception Networks. By using the XCovNet model, the previous study's problems are cautiously addressed and overhauled by achieving 99.76% accuracy, 99.89% specificity, 99.87% sensitivity, and 99.75% F1-score. To understand the underlying behavior of the proposed network, different tests are performed on different shuffle patterns. Thus, the proposed "XCovNet" can, in regions where test kits are limited, be used to help radiologists detect COVID-19 patients through ultrasound images in the current COVID-19 situation.

Exploring the role of texture features in deep convolutional neural networks: Insights from Portilla-Simoncelli statistics

It is well-understood that the performance of Deep Convolutional Neural Networks (DCNNs) in image recognition tasks is influenced not only by shape but also by texture information. Despite this, understanding the internal representations of DCNNs remains a challenging task. This study employs a simplified version of the Portilla-Simoncelli Statistics, termed “minPS,” to explore how texture information is represented in a pre-trained VGG network. Using minPS features extracted from texture images, we perform a sparse regression on the activations across various channels in VGG layers. Our findings reveal that channels in the early to middle layers of the VGG network can be effectively described by minPS features. Additionally, we observe that the explanatory power of minPS sub-groups evolves as one ascends the network hierarchy. Specifically, sub-groups termed Linear Cross Scale (LCS) and Energy Cross Scale (ECS) exhibit weak explanatory power for VGG channels. To investigate the relationship further, we compare the original texture images with their synthesized counterparts, generated using VGG, in terms of minPS features. Our results indicate that the absence of certain minPS features suggests their non-utilization in VGG’s internal representations.

CPD-CCNN: classification of pepper disease using a concatenation of convolutional neural network models

Agricultural products are vital to the sustainability of the economies of developing countries. Most developing countries’ economies such as Ethiopia heavily rely on agriculture. On a global scale, the pepper crop is one of the most important agricultural products in terms of human food security. However, it is susceptible to a variety of diseases which include blight leaf disease, gray leaf spot, common rust, fruit rot disease, powdery mildew symptoms on pepper leaf, and other related diseases that are all common today. Currently, more than 34 different pepper diseases have been discovered, resulting in a 33% average yield loss in pepper cultivation. Conventionally, farmers detect the disease using visual observation but this has its own demerits as it is usually not accurate and usually time consuming. In the past, a number of researchers have presented various methods for classifying pepper plant disease, especially using image processing and deep learning techniques. However, earlier studies have shown that binary classification requires improvement as some classes were more challenging to identify than others. In this study, we propose a concatenated neural network of the extracted features of VGG16 and AlexNet networks to develop a pepper disease classification model using fully connected layers. The development of the proposed concatenated CNN model includes steps such as dataset collection, image preprocessing, noise removal, segmentation, feature extraction, and classification. Finally, the proposed concatenated CNN model was evaluated, providing a training classification accuracy of 100%, validation accuracy of 97.29%, and testing accuracy of 95.82%. In general, it can be concluded from the findings of the study that the proposed concatenated model is suitable for identifying pepper leaf and fruit diseases from digital images of pepper.

Recognition of shear and tension signals based on acoustic emission parameters and waveform using machine learning methods

Acoustic emission (AE) technology is widely used to monitor the damage evolution of rock. Identifying AE signals is crucial to reveal the rock cracking mechanism. The current tension and shear signal discrimination methods primarily rely on AE parameters. However, the differences between AE parameters and waveform images used as inputs require further exploration. Acoustic emission data from Brazilian splitting and direct shear tests of red sandstone were collected. A decision tree was applied to classify the signals based on the AE parameters: Rise time (RT), Counts, Amplitude, Peak frequency (PF), and absolute energy. Three typical CNN-based networks (InceptionV3, ResNet50, and VGG19) were used to classify the AE signals using images of waveforms, reconfigured waveforms, and spectrograms generated by wavelet packets. The results showed that using AE parameters as input to distinguish tensile and shear signals yields an accuracy of up to 88%, increasing recognition accuracy when using PF as input. In addition, increasing the number of input parameters to the decision tree can improve average recognition accuracy, with single, double, and three parameters being 65%, 78%, and 82%, respectively. However, when the number of parameters exceeds three, the accuracy improvement is minimal, with four and five parameters being 84% and 86%, respectively. The CNN-based networks outperform the decision tree in terms of recognition accuracy. It is recommended that the VGG19 network and waveform are adopted as input to identify the tension and shear AE signals. In contrast, using the spectrogram as input enhances the stability of the model, particularly for the validation set, although the improvement is limited.

Explainable Artificial Intelligence Method (ParaNet+) Localises Abnormal Parathyroid Glands in Scintigraphic Scans of Patients with Primary Hyperparathyroidism

The pre-operative localisation of abnormal parathyroid glands (PG) in parathyroid scintigraphy is essential for suggesting treatment and assisting surgery. Human experts examine the scintigraphic image outputs. An assisting diagnostic framework for localisation reduces the workload of physicians and can serve educational purposes. Former studies from the authors suggested a successful deep learning model, but it produced many false positives. Between 2010 and 2020, 648 participants were enrolled in the Department of Nuclear Medicine of the University Hospital of Patras, Greece. An innovative modification of the well-known VGG19 network (ParaNet+) is proposed to classify scintigraphic images into normal and abnormal classes. The Grad-CAM++ algorithm is applied to localise the abnormal PGs. An external dataset of 100 patients imaged at the same department who underwent parathyroidectomy in 2021 and 2022 was used for evaluation. ParaNet+ agreed with the human readers, showing 0.9861 on a patient-level and 0.8831 on a PG-level basis under a 10-fold cross-validation on the training set of 648 participants. Regarding the external dataset, the experts identified 93 of 100 abnormal patient cases and 99 of 118 surgically excised abnormal PGs. The human-reader false-positive rate (FPR) was 10% on a PG basis. ParaNet+ identified 99/100 abnormal cases and 103/118 PGs, with an 11.2% FPR. The model achieved higher sensitivity on both patient and PG bases than the human reader (99.0% vs. 93% and 87.3% vs. 83.9%, respectively), with comparable FPRs. Deep learning can assist in detecting and localising abnormal PGs in scintigraphic scans of patients with primary hyperparathyroidism and can be adapted to the everyday routine.

Phenotype Analysis of Arabidopsis thaliana Based on Optimized Multi-Task Learning

Deep learning techniques play an important role in plant phenotype research, due to their powerful data processing and modeling capabilities. Multi-task learning has been researched for plant phenotype analysis, which can combine different plant traits and allow for a consideration of correlations between multiple phenotypic features for more comprehensive analysis. In this paper, an intelligent and optimized multi-task learning method for the phenotypic analysis of Arabidopsis thaliana is proposed and studied. Based on the VGG16 network, hard parameter sharing and task-dependent uncertainty are used to weight the loss function of each task, allowing parameters associated with genotype classification, leaf number counting, and leaf area prediction tasks to be learned jointly. The experiments were conducted on the Arabidopsis thaliana dataset, and the proposed model achieved weighted classification accuracy, precision, and Fw scores of 96.88%, 97.50%, and 96.74%, respectively. Furthermore, the coefficient of determination R2 values in the leaf number and leaf area regression tasks reached 0.7944 and 0.9787, respectively.

An approach to privacy-aware image analysis on edge devices using CNNs

Abstract This paper presents our approach to a privacy preserving person detection algorithm on edge computing devices. It utilizes a commonly used neural network architecture (VGG16) to encode an image by passing it through a fixed number of layers of the network on the edge device. The resulting feature vector can then be transmitted over the network to a more powerful computer (e.g. in the cloud) to be passed through the remaining layers of VGG16. As a result, the transmitted feature vector is only an abstract representation of the image. However, our research has shown that it is in fact possible to reconstruct most images by their respecting feature vectors if a potential attacker can tap into the network transmission by using a decoder network. Our approach to mitigate that risk is to preprocess the transmitted feature vector randomly by different manipulation methods. We used methods like mean filter or random null value insertion to manipulate the feature vector before it gets transmitted and show that those methods are able to counter the reconstruction capabilities of a decoder network while still preserving the capabilities of the original VGG16 network. The actual VGG16 network was used for object detection.We used the faces in the wild dataset and utilized two different approaches to confirm our approach. First, we used the VGGFace-Network on the decoded images and tried to let it identify people from the former mentioned dataset. The second approach used a group of 26 participants who had to match the decoder image to one of five images. In our experiments, we found different combinations of manipulation and number of layers on the edge device to preserve the detection capability of VGG16 (e.g. object detection) while preventing VGGFace and the participating group from identifying the shown people.

Human Action Recognition Based on Skeleton Information and Multi-Feature Fusion

Action assessment and feedback can effectively assist fitness practitioners in improving exercise benefits. In this paper, we address key challenges in human action recognition and assessment by proposing innovative methods that enhance performance while reducing computational complexity. Firstly, we present Oct-MobileNet, a lightweight backbone network, to overcome the limitations of the traditional OpenPose algorithm’s VGG19 network, which exhibits a large parameter size and high device requirements. Oct-MobileNet employs octave convolution and attention mechanisms to improve the extraction of high-frequency features from the human body contour, resulting in enhanced accuracy with reduced model computational burden. Furthermore, we introduce a novel approach for action recognition that combines skeleton-based information and multiple feature fusion. By extracting spatial geometric and temporal characteristics from actions, we employ a sliding window algorithm to integrate these features. Experimental results show the effectiveness of our approach, demonstrating its ability to accurately recognize and classify various human actions. Additionally, we address the evaluation of traditional fitness exercises, specifically focusing on the BaDunJin movements. We propose a multimodal information-based assessment method that combines pose detection and keypoint analysis. Label sequences are obtained through a pose detector and each frame’s keypoint coordinates are represented as pose vectors. Leveraging multimodal information, including label sequences and pose vectors, we explore action similarity and perform quantitative evaluations to help exercisers assess the quality of their exercise performance.

Dual-branch counting method for dense crowd based on self-attention mechanism

A dense crowd counting method based on self-attention mechanism with dual-branch fusion network is proposed in this paper. Our method aims to address the problems of large variations in head scales and complex backgrounds in dense crowd images. This method combines the CNN and Transformer network frameworks and consists of shallow feature extraction network, dual-branch fusion network, and deep feature extraction network. The VGG16 network is employed by the shallow feature extraction network to extract low-level features. A multi-scale CNN branch and a Transformer branch built on an improved self-attention module make up the dual-branch fusion network, which collects local and global information on crowd areas, respectively. The Transformer network, which is based on a mixed attention module, is employed by the deep feature extraction network to further separate complicated backgrounds and concentrate on crowd areas. Both counting-level weakly supervised and location-level fully supervised methods are employed in the experiments. On four widely used datasets, the results demonstrate that the proposed method outperforms the most recent research. Our method has a higher counting accuracy with low parameter volumes and a counting accuracy of 89.1% under full supervision when compared to existing weakly supervised methods. The results of the experiments demonstrate that the method has excellent crowd counting performance and can accurately count in high-density and high-occlusion scenes.

Intra-nucleus mosaic pattern (InMop) and whole-cell Haralick combined-descriptor for identifying and characterizing acute leukemia blasts on single cell peripheral blood images.

Acute leukemia is usually diagnosed when a test of peripheral blood shows at least 20% of abnormal immature cells (blasts), a figure even lower in case of recurrent cytogenetic abnormalities. Blast identification is crucial for white blood cell (WBC) counting, which depends on both identifying the cell type and characterizing the cellular morphology, processes susceptible of inter- and intraobserver variability. The present work introduces an image combined-descriptor to detect blasts and determine their probable lineage. This strategy uses an intra-nucleus mosaic pattern (InMop) descriptor that captures subtle nuclei differences within WBCs, and Haralick's statistics which quantify the local structure of both nucleus and cytoplasm. The InMop captures WBC inner-nucleus structure by applying a multiscale Shearlet decomposition over a repetitive pattern (mosaic) of automatically-segmented nuclei. As a complement, Haralick's statistics characterize the local structure of the whole cell from an intensity co-occurrence matrix representation. Both InMoP and Haralick-based descriptors are calculated using the b-channel from Lab color-space. The combined-descriptor is assessed by differentiating blasts from nonleukemic cells with support vector machine (SVM) classifiers and different transformation kernels, in two public and independent databases. The first database-D1 (n = 260) is composed of healthy and acute lymphoid leukemia (ALL) single cell images, and second database-D2 contains acute myeloid leukemia (AML) blasts (n = 3294) and nonblast (n = 15,071) cell images. In a first experiment, blasts versus nonblast differentiation is performed by training with a subset of D2 (n = 6588) and testing in D1 (n = 260), obtaining a training AUC of 0.991 ± 0.002 and AUC = 0.782 for the independent validation. A second experiment automatically differentiates AML blasts (260 images from D2) from ALL blasts (260 images from D1), with an AUC of 0.93. In a third experiment, state-of-the-art strategies, VGG16 and RESNEXT convolutional neural networks (CNN), separate blast from nonblast cells in both databases. The VGG16 showed an AUC of 0.673 and the RESNEXT of 0.75. Reported metrics for all the experiments are area under the ROC curve (AUC), accuracy and F1-score.

A Bio-Inspired Method For Breast Histopathology Image Classification Using Transfer Learning

Breast cancer is one of the deadly cancer among the female population, and still a developing area of research in the field of medical imaging. The fatality rate is more in patients who are not early diagnosed and are given delayed treatment. Hence researchers are keeping their lot of efforts in developing breast cancer detection systems that could provide accurate diagnosis in the initial stages which are relied on medical imaging. Deep learning is offering key solutions to overcome many image classification tasks. Though deep learning techniques have extended their root to many medical fields even it suffers from the problem of lack of sufficient data. Convolutional Neural Networks are more preferred for medical image classification tasks. In this paper, we propose a transfer learning method that overcomes the issue of insufficient data. We use a familiar pre-trained network VGG-16 (Visual Geometric Group) + with Logistic Regression as a binary classifier. Since hyper-parameters of every CNN has a closer impact on the performance of the entire deep learning model, our method focus on optimizing hyper-parameters using particle swarm optimization which is a bio-inspired algorithm, The proposed model performs classification of Breast Histopathology images into benign and malignant images and produce better results. We use Break His Dataset to test our method and achieve an accuracy of around 96.9%. The experimental results show that this study has improved performance metrics when compared to other transfer learning methods.

Quad-Pol SAR Data Reconstruction from Dual-Pol SAR Mode Based on a Multiscale Feature Aggregation Network

Polarimetric synthetic aperture radar (PolSAR) is widely used in remote sensing applications due to its ability to obtain full-polarization information. Compared to the quad-pol SAR, the dual-pol SAR mode has a wider observation swath and is more common in most SAR systems. The goal of reconstructing quad-pol SAR data from the dual-pol SAR mode is to learn the contextual information of dual-pol SAR images and the relationships among polarimetric channels. This work is dedicated to addressing this issue, and a multiscale feature aggregation network has been established to achieve the reconstruction task. Firstly, multiscale spatial and polarimetric features are extracted from the dual-pol SAR images using the pretrained VGG16 network. Then, a group-attention module (GAM) is designed to progressively fuse the multiscale features extracted by different layers. The fused feature maps are interpolated and aggregated with dual-pol SAR images to form a compact feature representation, which integrates the high- and low-level information of the network. Finally, a three-layer convolutional neural network (CNN) with a 1 × 1 convolutional kernel is employed to establish the mapping relationship between the feature representation and polarimetric covariance matrices. To evaluate the quad-pol SAR data reconstruction performance, both polarimetric target decomposition and terrain classification are adopted. Experimental studies are conducted on the ALOS/PALSAR and UAVSAR datasets. The qualitative and quantitative experimental results demonstrate the superiority of the proposed method. The reconstructed quad-pol SAR data can better sense buildings’ double-bounce scattering changes before and after a disaster. Furthermore, the reconstructed quad-pol SAR data of the proposed method achieve a 97.08% classification accuracy, which is 1.25% higher than that of dual-pol SAR data.

A two-stage recognition method based on deep learning for sheep behavior

Recently, computer vision has been widely used in livestock behavior detection. In animal multi-behavior detection, there is a common problem of low recognition accuracy. To address this, a deep learning-based method for sheep-behavior detection is proposed in this paper. Six types of sheep behavior can be recognized by the method. Among these behaviors, standing, feeding, and lying are normal physiological activities. Attacking, biting, and climbing are disruptive behaviors that could result in ranch losses and require immediate concern from the breeder. The method consists of two stages: the detection stage and the classification stage. In the detection stage, a detection network is used to determine whether every sheep's behavior belongs to normal physiological activities or disruptive behaviors respectively. Based on a classic network, the multi-scale feature aggregation, attention mechanism and depthwise convolution module are imported to it, which makes the network trades off between detection accuracy and model size. These improvements make the network more suitable for sheep-behavior detection. In the classification stage, the VGG network is utilized to classify the behavior of each sheep specifically. Experimental results demonstrate that the two-stage method achieves desirable results in sheep-behavior recognition. In the detection stage, the mAP of the two types of behaviors exceeds 98%. In the classification stage, the classification accuracy of all behaviors exceeds 94%. At the same time, the memory of the detection model is less than 130 MB.

Deep learning enables the differentiation between early and late stages of hip avascular necrosis

ObjectivesTo develop a deep learning methodology that distinguishes early from late stages of avascular necrosis of the hip (AVN) to determine treatment decisions.MethodsThree convolutional neural networks (CNNs) VGG-16, Inception ResnetV2, InceptionV3 were trained with transfer learning (ImageNet) and finetuned with a retrospectively collected cohort of (n = 104) MRI examinations of AVN patients, to differentiate between early (ARCO 1–2) and late (ARCO 3–4) stages. A consensus CNN ensemble decision was recorded as the agreement of at least two CNNs. CNN and ensemble performance was benchmarked on an independent cohort of 49 patients from another country and was compared to the performance of two MSK radiologists. CNN performance was expressed with areas under the curve (AUC), the respective 95% confidence intervals (CIs) and precision, and recall and f1-scores. AUCs were compared with DeLong’s test.ResultsOn internal testing, Inception-ResnetV2 achieved the highest individual performance with an AUC of 99.7% (95%CI 99–100%), followed by InceptionV3 and VGG-16 with AUCs of 99.3% (95%CI 98.4–100%) and 97.3% (95%CI 95.5–99.2%) respectively. The CNN ensemble the same AUCs Inception ResnetV2. On external validation, model performance dropped with VGG-16 achieving the highest individual AUC of 78.9% (95%CI 51.6–79.6%) The best external performance was achieved by the model ensemble with an AUC of 85.5% (95%CI 72.2–93.9%). No significant difference was found between the CNN ensemble and expert MSK radiologists (p = 0.22 and 0.092 respectively).ConclusionAn externally validated CNN ensemble accurately distinguishes between the early and late stages of AVN and has comparable performance to expert MSK radiologists.Clinical relevance statementThis paper introduces the use of deep learning for the differentiation between early and late avascular necrosis of the hip, assisting in a complex clinical decision that can determine the choice between conservative and surgical treatment.Key Points• A convolutional neural network ensemble achieved excellent performance in distinguishing between early and late avascular necrosis.• The performance of the deep learning method was similar to the performance of expert readers.