Robust Scene Text Detection with Deep Feature Pyramid Network and CNN based NMS Model

Background Incidence rate of thyroid cancer is steadily increasing due to overdiagnosis and overtreatment. Thyroid ultrasound is commonly used to diagnose thyroid cancer. The aim of this study is to examine the accuracy of using deep convolutional neural network (DCNN) models to improve diagnosis of thyroid cancer by analyzing sonographic imaging data from clinical thyroid ultrasound. Methods A total of 131,731 sonographic images from 17,627 thyroid cancer patients and 180,668 sonographic images from 25,325 controls used as training set were obtained from Tianjin Cancer Hospital. Images from anatomical sites that did not have cancer according to location sign on the image were not included. All thyroid cancer patients and 13·2% of controls (51,255 images) were confirmed by pathological reports. DCNN is a specific type of neural network optimized for image recognition. We trained two DCNN models on the training set and subsequently evaluated the performance on one independent internal (Tianjin, 1,118 individuals) and two external (Jilin,154 individuals; Weihai, 1,420 individuals) validation sets. Individuals in the validation sets all have pathological examinations. We compared the specificity/sensitivity of DCNN models with the performance of six thyroid ultrasound radiologists on these three validation sets. Findings DCNN model achieved high performance in identifying thyroid cancer patients versus six experience radiologists: for Tianjin validation set, sensitivity was 92·2% versus 96·9% (95% CI 89·7% - 94·3% vs. 93·9% - 98·6%; p = 0·003), and specificity was 85·6% versus 59·4% (95% CI 82·4% - 88·4% vs. 53% - 65·6%; p < 0·0001); for Jilin validation set, sensitivity was 84·3% versus 92·9% (95% CI 73·6% - 91·9% vs. 84·1% - 97·6%; p = 0·05), and specificity was 86·9% versus 57·1% (95% CI 77·8% - 93·3% vs. 45·9% - 67·9%; p < 0·0001); for Weihai validation set, sensitivity was 84·5% versus 89% (95% CI 81·2% - 87·4% vs. 81·9% - 94%; p = 0·2), and specificity was 87·5% versus 68·6% (95% CI 85·1% - 89·6% vs. 60·7% - 75·8%; p < 0·0001). Interpretation DCNN models exhibited high accuracy, sensitivity, and specificity in identifying thyroid cancer patients at levels comparable to or higher than six experienced radiologists. Conferred by the high specificity of DCNN models, the rate of overdiagnosis and overtreatment of patients with thyroid cancer is expected to decrease. This supports future application of the deep learning models to clinical practice for thyroid cancer diagnosis. However, further validation of these DCNN models in prospective clinical trials is warranted. Funding The Program for Changjiang Scholars and Innovative Research Team in University in China (IRT_14R40), National Natural Science Foundation of China (31801117). Citation Format: Xiangchun Li, Sheng Zhang, Qiang Zhang, Xi Wei, Yi Pan, Jing Zhao, Xiaojie Xin, Xiaoqing Wang, Fan Yang, Jianxin Li, Meng Yang, Qinghua Wang, Xiangqian Zheng, Yanhui Zhao, Lun Zhang, Xudong Wang, Zhimin Zheng, Christopher T. Whitlow, Metin N. Gurcan, Boris Pasche, Ming Gao, Wei Zhang, Kexin Chen. Diagnosis of thyroid cancer using deep convolutional neural network models applied to sonographic images from clinical ultrasound exams [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 1394.

The aim of the present study was to explore the diagnostic value of a deep convolutional neural network (DCNN) model for the diagnosis of pulmonary nodules in adolescent and young adult patients with osteosarcoma. For the present study, 675 chest CT images were retrospectively collected from 109 patients with clinically confirmed osteosarcoma who underwent chest CT examination at Hangzhou Third People's Hospital (Hangzhou, China) from March 2011 to February 2022. CT images were then evaluated using the DCNN and manual models. Subsequently, pulmonary nodules of osteosarcoma were divided into calcified nodules, solid nodules, partially solid nodules and ground glass nodules using the DCNN model. Those patients with osteosarcoma who were diagnosed and treated were followed up to observe dynamic changes in the pulmonary nodules. A total of 3,087 nodules were detected, while 278 nodules were missed compared with those determined using the reference standard given by the consensus of three Experienced radiologists., which was analyzed by two diagnostic radiologists. In the manual model group, 2,442 nodules were detected, while 657 nodules were missed. The DCNN model showed significantly higher sensitivity and specificity compared with the manual model (sensitivity, 0.923 vs. 0.908; specificity, 0.552 vs. 0.351; P<0.05). In addition, the DCNN model yielded an area under the curve (AUC) value of 0.795 [95% confidence interval (CI), 0.743-0.846], outperforming that of the manual model (AUC, 0.687; 95% CI, 0.629-0.732; P<0.05). The film reading time of the DCNN model was also significantly shorter compared with that of the manual model [mean ± standard deviation (SD); 173.25±24.10 vs. 328.32±22.72 sec; P<0.05)]. The AUC of calcified nodules, solid nodules, partially solid nodules and ground glass nodules was calculated to be 0.766, 0.771, 0.761 and 0.796, respectively, using the DCNN model. Using this model, the majority of the pulmonary nodules were detected in patients with osteosarcoma at the initial diagnosis (69/109, 62.3%), and the majority of these were found with multiple pulmonary nodules instead of a single nodule (71/109, 65.1% vs. 38/109, 34.9%). These data suggest that, compared with the manual model, the DCNN model proved to be beneficial for the detection of pulmonary nodules in adolescent and young adult patients with osteosarcoma, which may reduce the time of artificial radiograph reading. In conclusion, the proposed DCNN model, developed using data from 675 chest CT images retrospectively collected from 109 patients with clinically confirmed osteosarcoma, may be used as an effective tool to evaluate pulmonary nodules in patients with osteosarcoma.

Deep learning-based computed tomography applied to the diagnosis of rib fractures

BackgroundSurgical process modeling automatically identifies surgical phases, and further improvement in recognition accuracy is expected with deep learning. Surgical tool or time series information has been used to improve the recognition accuracy of a model. However, it is difficult to collect this information continuously intraoperatively. The present study aimed to develop a deep convolution neural network (CNN) model that correctly identifies the surgical phase during laparoscopic cholecystectomy (LC).MethodsWe divided LC into six surgical phases (P1–P6) and one redundant phase (P0). We prepared 115 LC videos and converted them to image frames at 3 fps. Three experienced doctors labeled the surgical phases in all image frames. Our deep CNN model was trained with 106 of the 115 annotation datasets and was evaluated with the remaining datasets. By depending on both the prediction probability and frequency for a certain period, we aimed for highly accurate surgical phase recognition in the operation room.ResultsNine full LC videos were converted into image frames and were fed to our deep CNN model. The average accuracy, precision, and recall were 0.970, 0.855, and 0.863, respectively.ConclusionThe deep learning CNN model in this study successfully identified both the six surgical phases and the redundant phase, P0, which may increase the versatility of the surgical process recognition model for clinical use. We believe that this model can be used in artificial intelligence for medical devices. The degree of recognition accuracy is expected to improve with developments in advanced deep learning algorithms.

Electricity theft is harmful to power grids. Integrating information flows with energy flows, smart grids can help to solve the problem of electricity theft owning to the availability of massive data generated from smart grids. The data analysis on the data of smart grids is helpful in detecting electricity theft because of the abnormal electricity consumption pattern of energy thieves. However, the existing methods have poor detection accuracy of electricity theft since most of them were conducted on one-dimensional (1-D) electricity consumption data and failed to capture the periodicity of electricity consumption. In this paper, we originally propose a novel electricity-theft detection method based on wide and deep convolutional neural networks (CNN) model to address the above concerns. In particular, wide and deep CNN model consists of two components: the wide component and the deep CNN component. The deep CNN component can accurately identify the nonperiodicity of electricity theft and the periodicity of normal electricity usage based on 2-D electricity consumption data. Meanwhile, the wide component can capture the global features of 1-D electricity consumption data. As a result, wide and deep CNN model can achieve the excellent performance in electricity-theft detection. Extensive experiments based on realistic dataset show that wide and deep CNN model outperforms other existing methods.

Land cover and its dynamic information is the basis for characterizing surface conditions, supporting land resource management and optimization, and assessing the impacts of climate change and human activities. In land cover information extraction, the traditional convolutional neural network (CNN) method has several problems, such as the inability to be applied to multispectral and hyperspectral satellite imagery, the weak generalization ability of the model and the difficulty of automating the construction of a training database. To solve these problems, this study proposes a new type of deep convolutional neural network based on Landsat-8 Operational Land Imager (OLI) imagery. The network integrates cascaded cross-channel parametric pooling and average pooling layer, applies a hierarchical sampling strategy to realize the automatic construction of the training dataset, determines the technical scheme of model-related parameters, and finally performs the automatic classification of remote sensing images. This study used the new type of deep convolutional neural network to extract land cover information from Qinhuangdao City, Hebei Province, and compared the experimental results with those obtained by traditional methods. The results show that: (1) The proposed deep convolutional neural network (DCNN) model can automatically construct the training dataset and classify images. This model performs the classification of multispectral and hyperspectral satellite images using deep neural networks, which improves the generalization ability of the model and simplifies the application of the model. (2) The proposed DCNN model provides the best classification results in the Qinhuangdao area. The overall accuracy of the land cover data obtained is 82.0%, and the kappa coefficient is 0.76. The overall accuracy is improved by 5% and 14% compared to the support vector machine method and the maximum likelihood classification method, respectively.

Diagnosing burns in humans has become critical, as early identification can save lives. The manual process of burn diagnosis is time-consuming and complex, even for experienced doctors. Machine learning (ML) and deep convolutional neural network (CNN) models have emerged as the standard for medical image diagnosis. The ML-based approach typically requires handcrafted features for training, which may result in suboptimal performance. Conversely, DL-based methods automatically extract features, but designing a robust model is challenging. Additionally, shallow DL methods lack long-range feature dependency, decreasing efficiency in various applications. We implemented several deep CNN models, ResNeXt, VGG16, and AlexNet, for human burn diagnosis. The results obtained from these models were found to be less reliable since shallow deep CNN models need improved attention modules to preserve the feature dependencies. Therefore, in the proposed study, the feature map is divided into several categories, and the channel dependencies between any two channel mappings within a given class are highlighted. A spatial attention map is built by considering the links between features and their locations. Our attention-based model BuRnGANeXt50 kernel and convolutional layers are also optimized for human burn diagnosis. The earlier study classified the burn based on depth of graft and non-graft. We first classified the burn based on the degree. Subsequently, it is classified into graft and non-graft. Furthermore, the proposed model performance is evaluated on Burns_BIP_US_database. The sensitivity of the BuRnGANeXt50 is 97.22% and 99.14%, respectively, for classifying burns based on degree and depth. This model may be used for quick screening of burn patients and can be executed in the cloud or on a local machine. The code of the proposed method can be accessed at https://github.com/dhirujis02/Journal.git for the sake of reproducibility.

Recent studies have demonstrated that a well designed deep convolutional neural network (CNN) model achieves competitive performances on detecting the presence of secret message in digital images, compared with the classical rich model based steganalysis. In this paper, we propose to investigate a category of very deep CNN model–the deep residual network (DRN), for steganalysis. DRN is suitable for steganalysis from two aspects. For the first, the DRN model usually contains a large number of network layers, which proves to be effective to capture the complex statistics of digital images. For the second, DRN's residual learning (ResL) method actively strengthens the signal coming from secret messages, which is extremely beneficial for the discrimination between cover images and stego images. Comprehensive experiments on standard dataset show that the DRN model achieves very low detection error rates for the state of arts steganographic algorithms. It also outperforms the classical rich model method and several recently proposed CNN based methods.

The diabetic foot ulcer (DFU) is a severe complication that affects approximately 33% of diabetes patients globally, often leading to limb amputation if not detected early. This study introduces an automated approach for identifying and classifying DFU using transfer learning. DFU is typically categorized into ischemic and infection states, which are challenging to distinguish visually. We evaluate the effectiveness of pretrained Deep Convolutional Neural Network (DCNN) models for autonomous DFU detection. Seven models are compared: EfficientNetB0, DenseNet121, ResNet101, VGG16, MobileNetV2, InceptionV3, and InceptionResNetV2. Additionally, we propose E-DFu-Net, a novel model derived from existing architectures, designed to mitigate overfitting. Experimental results demonstrate that E-DFu-Net achieves remarkable performance, with 97% accuracy in ischemia classification and 92% in infection classification. This advancement enhances current methodologies and aids practitioners in effectively detecting DFU cases.

Automated image classification is an essential task of the computer vision field. The tagging of images into a set of predefined groups is referred to as image classification. The implementation of computer vision to automate image classification would be beneficial because manual image evaluation and identification can be time-consuming, particularly when there are many images of different classes. Deep learning approaches are proven to overperform existing machine learning techniques in many fields in recent years, and computer vision is one of the most notable examples. The very deep neural network (VDCNN) is a powerful deep learning model for image classification, and this paper examines it briefly using MNIST handwritten digit dataset. This dataset is used to prove the efficacy of very deep neural networks over other deep learning models. The proposed study aims to comprehend the very deep neural network architecture used to accomplish a handwritten digit recognition task. The feasibility of the proposed model is evaluated using mean accuracy, validation accuracy, and standard deviation. The study results of the very deep neural network model are compared to a convolutional neural network and convolutional neural network with batch normalization. According to the results of the comparison study, very deep neural networks achieve high accuracy of 99.1% for a handwritten dataset. The outcome of the proposed work is used to interpret how well a very deep neural network performs when compared to the other two models of deep neural networks. This proposed architecture may be used to automate the classification of handwritten digits dataset.

For the purpose of applying information technology to the creation of ink style painting, the algorithm of ink painting rendering based on the deep learning framework and convolutional neural network model is designed and improved. Firstly, the ink style rendering program is written in Python. Secondly, VGG under Caffe architecture and Illustration 2Vec models are transplanted to TensorFlow architecture, and the image is rendered in ink style based on deep learning framework and convolutional neural network model. Finally, based on Node.js, the server-side program for image ink style rendering is built. Among them, Express is adopted as the Web-side framework, and the front-end page effect is completed. The results show that the ink rendering logic program is applicable, and the expected purpose is achieved.

The automatic narration of a natural scene is an important trait in artificial intelligence that unites computer vision and natural language processing. Caption generation is a challenging task in scene understanding. Most of the state-of-the-art methods are using deep convolutional neural network models to extract visual features of the entire image, based on which the parallel structures between images and sentences are exploited using recurrent neural networks for image captioning. However, in such models, only visual features are exploited for caption generation. This work investigated that fusion of text available in an image can give more fined-grained captioning of a scene. In this paper, we have proposed a model which incorporates a deep convolutional neural network and long short-term memory to boost the accuracy of image captioning by fusing text feature available in an image with the visual features extracted in state-of-the-art methods. We have validated the effectiveness of the proposed model on the benchmark datasets (Flickr8k and Flickr30k). The experimental outcomes illustrate that the proposed model outperformed the state-of-the-art methods for image captioning.

Computer-aided diagnosis of ocular disorders like glaucoma can be effectively performed with retinal fundus images. Since the advent of machine learning and later deep learning techniques in medical image analysis, research focusing on automated and early detection of glaucoma has gained prominence. In this paper, we show how a deep Convolutional Neural Network model can be used in visualising the features related to Glaucoma that are consistent with (a subset of) those used by ophthalmologists for diagnosis. However, since the doctors do not necessarily depend entirely on these features but back them up with other important measurements such as visual field testing (HVF), any deep neural network such as CNN may not necessarily be expected to yield 100% accuracy. Further, other important parameters such as Intraocular Pressure (IOP), which may be built into a deep learning model, may not necessarily correlate well with the presence of Glaucoma. In this work, a deep CNN model has been trained and tested on a large number of high-quality fundus images, both normal and glaucomatous. We compare the results with transfer learning models such as VGG16, ResNet50, and MobileNetV2. On an average, we obtained an accuracy of 93.75% in identifying glaucoma, focusing only on the features of the fundus image.

Deep neural network (DNN) models typically have many hyperparameters that can be configured to achieve optimal performance on a particular dataset. Practitioners usually tune the hyperparameters of their DNN models by training a number of trial models with different configurations of the hyperparameters, to find the optimal hyperparameter configuration that maximizes the training accuracy or minimizes the training loss. As such hyperparameter tuning usually focuses on the model accuracy or the loss function, it is not clear and remains under-explored how the process impacts other performance properties of DNN models, such as inference latency and model size. On the other hand, standard DNN models are often large in size and computing-intensive, prohibiting them from being directly deployed in resource-bounded environments such as mobile devices and Internet of Things (IoT) devices. To tackle this problem, various model optimization techniques (e.g., pruning or quantization) are proposed to make DNN models smaller and less computing-intensive so that they are better suited for resource-bounded environments. However, it is neither clear how the model optimization techniques impact other performance properties of DNN models such as inference latency and battery consumption, nor how the model optimization techniques impact the effect of hyperparameter tuning (i.e., the compounding effect). Therefore, in this paper, we perform a comprehensive study on four representative and widely-adopted DNN models, i.e., CNN image classification , Resnet-50 , CNN text classification , and LSTM sentiment classification , to investigate how different DNN model hyperparameters affect the standard DNN models, as well as how the hyperparameter tuning combined with model optimization affect the optimized DNN models, in terms of various performance properties (e.g., inference latency or battery consumption). Our empirical results indicate that tuning specific hyperparameters has heterogeneous impact on the performance of DNN models across different models and different performance properties. In particular, although the top tuned DNN models usually have very similar accuracy, they may have significantly different performance in terms of other aspects (e.g., inference latency). We also observe that model optimization has a confounding effect on the impact of hyperparameters on DNN model performance. For example, two sets of hyperparameters may result in standard models with similar performance but their performance may become significantly different after they are optimized and deployed on the mobile device. Our findings highlight that practitioners can benefit from paying attention to a variety of performance properties and the confounding effect of model optimization when tuning and optimizing their DNN models.

Scene text detection is a crucial step in end-to-end scene text recognition, a greatly challenging problem in computer vision. This paper proposes a novel scene text detection method that involves superpixel-based stroke feature transform (SSFT) and deep learning based region classification (DLRC). The SSFT is developed for candidate character region (CCR) extraction, which consists in partitioning an input image into several regions via superpixel-based clustering, removing most regions based on predefined criteria satisfied by the characters, and refining the remaining regions to obtain CCRs by computing a stroke width map. The character regions are identified from the CCRs using DLRC, in which several hand-crafted low-level features, i.e., color, texture, and geometric features, and some deep convolution neural network (CNN) based high-level features are first extracted from the regions, and then these features are fused by using two fully connected networks (FCNs) for region classification. In the DLRC step, the deep feature extraction CNN and the feature fusion FCNs are jointly trained. Next, the extracted character regions are merged to form candidate text regions, from which the final scene texts are detected. The proposed method is evaluated on three publicly available datasets: ICDAR2011, ICDAR2013, and street view text. It achieves F -measures of 0.876, 0.885, and 0.631, respectively, which demonstrate the effectiveness of the proposed scene text detection method.