Multi-class Convolutional Neural Network Research Articles

Multiclass Convolutional Neural Networks for Atrial Fibrillation Classification.

Atrial fibrillation (AF) is a common supraventricular arrhythmia. Its automatic identification by standard 12-lead electrocardiography (ECG) is still challenging. Recently, deep learning provided new instruments able to mimic the diagnostic ability of clinicians but only in case of binary classification (AF vs. normal sinus rhythm-NSR). However, binary classification is far from the real scenarios, where AF has to be discriminated also from several other physiological and pathological conditions. The aim of this work is to present a new AF multiclass classifier based on a convolutional neural network (CNN), able to discriminate AF from NSR, premature atrial contraction (PAC) and premature ventricular contraction (PVC). Overall, 2796 12-lead ECG recordings were selected from the open-source "PhysioNet/Computing in Cardiology Challenge 2021" database, to construct a dataset constituted by four balanced classes, namely AF class, PAC class, PVC class, and NSR class. Each lead of each ECG recording was decomposed into spectrogram by continuous wavelet transform and saved as 2D grayscale images, used to feed a 6-layers CNN. Considering the same CNN architecture, a multiclass classifiers (all classes) and three binary classifiers (AF class, PAC class, and PVC class vs. NSR class) were created and validated by a stratified shuffle split cross-validation of 10 splits. Performance was quantified in terms of area under the curve (AUC) of the receiver operating characteristic. Multiclass classifier performance was high (AF class: 96.6%; PAC class: 95.3%; PVC class: 92.8%; NSR class: 97.4%) and preferable to binary classifiers. Thus, our CNN AF multiclass classifier proved to be an efficient tool for AF discrimination from physiological and pathological confounders. Clinical Relevance-Our CNN AF multiclass classifier proved to be suitable for AF discrimination in real scenarios.

PromoterLCNN: A Light CNN-Based Promoter Prediction and Classification Model.

Promoter identification is a fundamental step in understanding bacterial gene regulation mechanisms. However, accurate and fast classification of bacterial promoters continues to be challenging. New methods based on deep convolutional networks have been applied to identify and classify bacterial promoters recognized by sigma () factors and RNA polymerase subunits which increase affinity to specific DNA sequences to modulate transcription and respond to nutritional or environmental changes. This work presents a new multiclass promoter prediction model by using convolutional neural networks (CNNs), denoted as PromoterLCNN, which classifies Escherichia coli promoters into subclasses , , , , , and . We present a light, fast, and simple two-stage multiclass CNN architecture for promoter identification and classification. Training and testing were performed on a benchmark dataset, part of RegulonDB. Comparative performance of PromoterLCNN against other CNN-based classifiers using four parameters (Acc, Sn, Sp, MCC) resulted in similar or better performance than those that commonly use cascade architecture, reducing time by approximately 30–90% for training, prediction, and hyperparameter optimization without compromising classification quality.

Convolution neural network model for fundus photograph quality assessment

<span>The excellent quality of color fundus photograph is crucial for the ophthalmologist to process the correct diagnosis and for convolutional neural network (CNN) models to optimize output classification. As a result of main causes as acquire devises efficiency and experience of a physician most fundus photographs can have uneven illuminance, blur, and bad contrast, in addition to micro-features of retinal diseases, which need to force their contrast. Fundus photograph quality assessment method is proposed to find out the perfect enhanced color fundus Technique in fundoscopy photographs-based CNN model. Five photograph quality measurements, in addition to five CNN metrics, were used as standard in this study. In this research innovative approach combining photograph quality measurement and CNN metrics analysis is proposed to find out the best enhance method that is set for the multiclass CNN model. The contrast enhancement techniques are evaluated using 267 color fundus photographs divided into three retina diseases cases were downloaded from the open-source database “FIGSHARE”. The study outcome showed that the presented system (single-CNN) can determine well the contrast enhancement method, as well as the low-quality fundus photograph then it can boost CNN metrics to achieve superior.</span>

Diagnosing and differentiating viral pneumonia and COVID-19 using X-ray images.

Coronavirus-caused diseases are common worldwide and might worsen both human health and the world economy. Most people may instantly encounter coronavirus in their life and may result in pneumonia. Nowadays, the world is fighting against the new coronavirus: COVID-19. The rate of increase is high, and the world got caught the disease unprepared. In most regions of the world, COVID-19 test is not possible due to the absence of the diagnostic kit, even if the kit exists, its false-negative (giving a negative result for a person infected with COVID-19) rate is high. Also, early detection of COVID-19 is crucial to keep its morbidity and mortality rates low. The symptoms of pneumonia are alike, and COVID-19 is no exception. The chest X-ray is the main reference in diagnosing pneumonia. Thus, the need for radiologists has been increased considerably not only to detect COVID-19 but also to identify other abnormalities it caused. Herein, a transfer learning-based multi-class convolutional neural network model was proposed for the automatic detection of pneumonia and also for differentiating non-COVID-19 pneumonia and COVID-19. The model that inputs chest X-ray images is capable of extracting radiographic patterns on chest X-ray images to turn into valuable information and monitor structural differences in the lungs caused by the diseases. The model was developed by two public datasets: Cohen dataset and Kermany dataset. The model achieves an average training accuracy of 0.9886, an average training recall of 0.9829, and an average training precision of 0.9837. Moreover, the average training false-positive and false-negative rates are 0.0085 and 0.0171, respectively. Conversely, the model’s test set metrics such as average accuracy, average recall, and average precision are 97.78%, 96.67%, and 96.67%, respectively. According to the simulation results, the proposed model is promising, can quickly and accurately classify chest images, and helps doctors as the second reader in their final decision.

Identifying and building on the current state of bioacoustics software

Bioacoustics is a powerful and increasingly commonly used tool for terrestrial and marine biological assessments. As the scale of bioacoustic data collection has increased, techniques for processing these data have diversified. However, with analysis methods rapidly evolving and dozens of analysis software packages already available, it is challenging to identify which software, if any, meets a particular researcher’s needs. We reviewed bioacoustics software to identify packages aimed at or used by bioacoustics researchers in ecology. We compiled descriptions of the function of 65 stable or actively developed software packages used for bioacoustics analyses. Of these, 59 were free or open-source packages. In addition, we developed free, open-source Python software, OpenSoundscape, that addresses gaps in available software. OpenSoundscape simplifies the process of creating flexible, scalable deep learning algorithms for bioacoustic analysis. It can be used to train binary or multiclass convolutional neural networks with any PyTorch-implemented model structure (e.g., ResNet50, Inception v3). Researchers can easily customize its spectrogram preprocessing and data augmentation routines to improve model performance. OpenSoundscape also includes modules to work with annotated acoustic data, apply additional signal processing algorithms, perform acoustic localization, and “open the black box” of deep learning using Grad-CAM.

A Convolutional Neural Network Bird Species Recognizer Built From Little Data by Iteratively Training, Detecting, and Labeling

Automatically detecting the calls of species of interest in audio recordings is a common but often challenging exercise in ecoacoustics. This challenge is increasingly being tackled with deep neural networks that generally require a rich set of training data. Often, the available training data might not be from the same geographical region as the study area and so may contain important differences. This mismatch in training and deployment datasets can impact the accuracy at deployment, mainly due to confusing sounds absent from the training data generating false positives, as well as some variation in call types. We have developed a multiclass convolutional neural network classifier for seven target bird species to track presence absence of these species over time in cotton growing regions. We started with no training data from cotton regions but we did have an unbalanced library of calls from other locations. Due to the relative scarcity of calls in recordings from cotton regions, manually scanning and labeling the recordings was prohibitively time consuming. In this paper we describe our process of overcoming this data mismatch to develop a recognizer that performs well on the cotton recordings for most classes. The recognizer was trained on recordings from outside the cotton regions and then applied to unlabeled cotton recordings. Based on the resulting outputs a verification set was chosen to be manually tagged and incorporated in the training set. By iterating this process, we were gradually able to build the training set of cotton audio examples. Through this process, we were able to increase the average class F1 score (the harmonic mean of precision and recall) of the recognizer on target recordings from 0.45 in the first iteration to 0.74.

A Non-Exclusive Multi-Class Convolutional Neural Network for the Classification of Functional Requirements in AUTOSAR Software Requirement Specification Text

Software Requirement Specification (SRS) describes a software system to be developed that captures the functional, non-functional, and technical aspects of the stakeholder’s requirements. Retrieval and extraction of software information from SRS are essential to the development of software product line (SPL). Albeit Natural Language Processing (NLP) techniques, such as information retrieval and standard machine learning, have been advocated in the recent past as a semi-automatic means of optimising requirements specifications, they have not been widely embraced. The complexity in the organization’s information makes requirement analysis intricately a challenging task. The interdependence of subsystems and within an organisation drives this complexity. A plain multi-class classification framework may not address this issue. Hence, this paper propounds an automated non-exclusive approach for classification of functional requirements from SRS, using a deep learning framework. Specifically, Word2Vec and FastText word embeddings are utilised for document representation for training a convolutional neural network (CNN). The study was carried out by the compilation of manually categorised relevant enterprise data (AUTomotive Open System ARchitecture (AUTOSAR)), which were also employed for model training. Over a convolutional neural network, the impact of data trained with Word2Vec and FastText word embeddings from SRS documentation were compared to pre-trained word embeddings models, available online.

Artificial intelligence for the assessment of bowel preparation

A reliable assessment of bowel preparation is important to ensure high-quality colonoscopy. Current bowel preparation scoring systems are limited by interobserver variability. This study aimed to demonstrate objective assessment of bowel preparation adequacy using an artificial intelligence (AI)/convolutional neural network (CNN) algorithm developed from colonoscopy videos. Two CNNs were developed using a training set of 73,304 images from 200 colonoscopies. First, a binary CNN was developed and trained to distinguish video frames that were appropriate versus inappropriate for scoring with the Boston Bowel Preparation Scale (BBPS). A second multiclass CNN was developed and trained on 26,950 appropriate frames that were expertly annotated with BBPS segment scores (0-3). We validated the algorithm using 252 10-second video clips that were assigned BBPS segment scores by 2 experts. The algorithm provided mean BBPS scores based on the algorithm (AI-BBPS) by calculating mean BBPS based on each frame's scoring. We maximized the algorithm's performance by choosing a dichotomized AI-BBPS score that closely matched dichotomized BBPS scores (ie, adequate vs inadequate). We tested the mean BBPS score based on the algorithm AI-BBPS against human rating using 30 independent 10-second video clips (test set 1) and 10 full withdrawal colonoscopy videos (test set 2). In the validation set, the algorithm demonstrated an area under the curve of .918 and accuracy of 85.3% for detection of inadequate bowel cleanliness. In test set 1, sensitivity for inadequate bowel preparation was 100% and agreement between raters and AI was 76.7% to 83.3%. In test set 2, sensitivity for inadequate bowel preparation for each segment was 100% and agreement between raters and AI was 68.9% to 89.7%. Agreement between raters alone versus raters and AI were similar (κ= .694 and .649, respectively). The algorithm assessment of bowel cleanliness as measured with the BBPS showed good performance and agreement with experts including full withdrawal colonoscopies.

Deep learning assisted cognitive diagnosis for the D-Riska application

In this article, we expose a system developed that extends the Acquired Brain Injury (ABI) diagnostic application known as D-Riska with an artificial intelligence module that supports the diagnosis of ABI enabling therapists to evaluate patients in an assisted way. The application is in charge of collecting the data of the diagnostic tests of the patients, and due to a multi-class Convolutional Neural Network classifier (CNN), it is capable of making predictions that facilitate the diagnosis and the final score obtained in the test by the patient. To find out the best solution to this problem, different classifiers are used to compare the performance of the proposed model based on various classification metrics. The proposed CNN classifier makes predictions with 93 % of Accuracy, 94 % of Precision, 91 %, of Recall and 92% of F1-Score.

Monitoring offshore oil pollution using multi-class convolutional neural networks

Oil and gas production operations are a major source of environmental pollution that expose people and habitats in many coastal communities around the world to adverse health effects. Detecting oil spills in a timely and precise manner can help improve the oil spill response process and channel required resources more effectively to affected regions. In this research, convolutional neural networks, a branch of artificial intelligence (AI), are trained on a visual dataset of oil spills containing images from different altitudes and geographical locations. In particular, a VGG16 model is adopted through transfer learning for oil spill classification (i.e., detecting if there is oil spill in an image) with an accuracy of 92%. Next, Mask R–CNN and PSPNet models are used for oil spill segmentation (i.e., pixel-level detection of oil spill boundaries) with a mean intersection over union (IoU) of 49% and 68%, respectively. Lastly, to determine if there is an oil rig or vessel in the vicinity of a detected oil spill and provide a holistic view of the oil spill surroundings, a YOLOv3 model is trained and used, yielding a maximum mean average precision (mAP) of ~71%. Findings of this research can improve the current practices of oil pollution cleanup and predictive maintenance, ultimately leading to more resilient and healthy coastal communities.

Development of A Computer Aided Real-Time Interpretation System for Indigenous Sign Language in Nigeria Using Convolutional Neural Network

Sign language is the primary method of communication adopted by deaf and hearing-impaired individuals. The indigenous sign language in Nigeria is one area receiving growing interest, with the major challenge faced is communication between signers and non-signers. Recent advancements in computer vision and deep learning neural networks (DLNN) have led to the exploration of necessary technological concepts towards tackling existing challenges. One area with extensive impact from the use of DLNN is the interpretation of hand signs. This study presents an interpretation system for the indigenous sign language in Nigeria. The methodology comprises three key phases: dataset creation, computer vision techniques, and deep learning model development. A multi-class Convolutional Neural Network (CNN) is designed to train and interpret the indigenous signs in Nigeria. The model is evaluated using a custom-built dataset of some selected indigenous words comprising of 15000 image samples. The experimental outcome shows excellent performance from the interpretation system, with accuracy attaining 95.67%.

Ensemble Method of Convolutional Neural Networks with Directed Acyclic Graph Using Dermoscopic Images: Melanoma Detection Application.

The early detection of melanoma is the most efficient way to reduce its mortality rate. Dermatologists achieve this task with the help of dermoscopy, a non-invasive tool allowing the visualization of patterns of skin lesions. Computer-aided diagnosis (CAD) systems developed on dermoscopic images are needed to assist dermatologists. These systems rely mainly on multiclass classification approaches. However, the multiclass classification of skin lesions by an automated system remains a challenging task. Decomposing a multiclass problem into a binary problem can reduce the complexity of the initial problem and increase the overall performance. This paper proposes a CAD system to classify dermoscopic images into three diagnosis classes: melanoma, nevi, and seborrheic keratosis. We introduce a novel ensemble scheme of convolutional neural networks (CNNs), inspired by decomposition and ensemble methods, to improve the performance of the CAD system. Unlike conventional ensemble methods, we use a directed acyclic graph to aggregate binary CNNs for the melanoma detection task. On the ISIC 2018 public dataset, our method achieves the best balanced accuracy (76.6%) among multiclass CNNs, an ensemble of multiclass CNNs with classical aggregation methods, and other related works. Our results reveal that the directed acyclic graph is a meaningful approach to develop a reliable and robust automated diagnosis system for the multiclass classification of dermoscopic images.

Collective human intelligence outperforms artificial intelligence in a skin lesion classification task.

Convolutional neural networks (CNN) enable accurate diagnosis of medical images and perform on or above the level of individual physicians. Recently, collective human intelligence (CoHI) was shown to exceed the diagnostic accuracy of individuals. Thus, diagnostic performance of CoHI (120 dermatologists) versus individual dermatologists versus two state-of-the-art CNN was investigated. Cross-sectional reader study with presentation of 30 clinical cases to 120 dermatologists. Six diagnoses were offered and votes collected via remote voting devices (quizzbox®, Quizzbox Solutions GmbH, Stuttgart, Germany). Dermatoscopic images were classified by a binary and multiclass CNN (FotoFinder Systems GmbH, Bad Birnbach, Germany). Three sets of diagnostic classifications were scored against ground truth: (1) CoHI, (2) individual dermatologists, and (3) CNN. CoHI attained a significantly higher accuracy [95% confidence interval] (80.0% [62.7%-90.5%]) than individual dermatologists (75.7% [73.8%-77.5%]) and CNN (70.0% [52.1%-83.3%]; all P<0.001) in binary classifications. Moreover, CoHI achieved a higher sensitivity (82.4% [59.0%-93.8%]) and specificity (76.9% [49.7%-91.8%]) than individual dermatologists (sensitivity 77.8% [75.3%-80.2%], specificity 73.0% [70.6%-75.4%]) and CNN (sensitivity 70.6% [46.9%-86.7%], specificity 69.2% [42.4%-87.3%]). The diagnostic accuracy of CoHI was superior to that of individual dermatologists (P < 0.001) in multiclass evaluation, with the accuracy of the latter comparable to multiclass CNN. Our analysis revealed that the majority vote of an interconnected group of dermatologists (CoHI) outperformed individuals and CNN in a demanding skin lesion classification task.

A combined finite element and hierarchical Deep learning approach for structural health monitoring: Test on a pin-joint composite truss structure

Structural Health Monitoring (SHM) is an emerging field of engineering with a wide range of applications. The most common SHM strategies operate on structural responses through vibration measurements and focus on training mathematical classifiers which are used after to identify damage in unknown responses. Classifiers may additionally locate damage when adequate labeled damaged data is available. In the present work, a novel SHM method is presented where labeled damaged data is generated through FE models for a pin-joint composite truss structure employing a model-based approach for the problem of data acquisition. The truss is made of carbon fiber reinforced polymer (CFRP) members joint on aluminum connections forming a complex and large FE problem. A Deep Learning (DL) Convolutional Neural Network (CNN) classifier is trained on the FE generated vibration data combined with a hierarchical multiple damage identification and location scheme. The numerically trained CNN is after validated on experimental statuses of the truss in both damage detection and location, proving to be robust and accurate for the considered test case. The potential of hierarchical CNNs with FE based SHM data for multiple damages is investigated in this work and a comparison is given between hierarchical and direct multiclass CNNs. The large performance gains of the former are proven for the studied experimental case highlighting also the importance of SHM system architectures with CNNs.

Heart Sound Multiclass Analysis Based on Raw Data and Convolutional Neural Network

In recent years, automatic diagnosis of the state of health of the heart by employing the phonocardiogram (PCG) has achieved remarkable success. This letter proposes a low-complexity automated solution based on the direct application of a multiclass convolutional neural network to PCG signals for the purpose of recognizing and classifying heart disease. PCG signals are fed in to the neural network, bypassing transformations from the time domain to that of frequencies (for example, mel-frequency cepstral coefficients (MFCC), Wavelet, etc.). Applying a recurrence filter in the postprocessing phase, the proposed method allows us to increase the performance in a range from 90 to 100%, with an analysis time of about 34 s. This letter also proposes an analysis of the robustness of the proposed technique to environmental office noise and to real-time PCG sequences.

Deep learning-based, computer-aided classifier developed with dermoscopic images shows comparable performance to 164 dermatologists in cutaneous disease diagnosis in the Chinese population.

BackgroundDiagnoses of Skin diseases are frequently delayed in China due to lack of dermatologists. A deep learning-based diagnosis supporting system can facilitate pre-screening patients to prioritize dermatologists’ efforts. We aimed to evaluate the classification sensitivity and specificity of deep learning models to classify skin tumors and psoriasis for Chinese population with a modest number of dermoscopic images.MethodsWe developed a convolutional neural network (CNN) based on two datasets from a consecutive series of patients who underwent the dermoscopy in the clinic of the Department of Dermatology, Peking Union Medical College Hospital, between 2016 and 2018, prospectively. In order to evaluate the feasibility of the algorithm, we used two datasets. Dataset I consisted of 7192 dermoscopic images for a multi-class model to differentiate three most common skin tumors and other diseases. Dataset II consisted of 3115 dermoscopic images for a two-class model to classify psoriasis from other inflammatory diseases. We compared the performance of CNN with 164 dermatologists in a reader study with 130 dermoscopic images. The experts’ consensus was used as the reference standard except for the cases of basal cell carcinoma (BCC), which were all confirmed by histopathology.ResultsThe accuracies of multi-class and two-class models were 81.49% ± 0.88% and 77.02% ± 1.81%, respectively. In the reader study, for the multi-class tasks, the diagnosis sensitivity and specificity of 164 dermatologists were 0.770 and 0.962 for BCC, 0.807 and 0.897 for melanocytic nevus, 0.624 and 0.976 for seborrheic keratosis, 0.939 and 0.875 for the “others” group, respectively; the diagnosis sensitivity and specificity of multi-class CNN were 0.800 and 1.000 for BCC, 0.800 and 0.840 for melanocytic nevus, 0.850 and 0.940 for seborrheic keratosis, 0.750 and 0.940 for the “others” group, respectively. For the two-class tasks, the sensitivity and specificity of dermatologists and CNN for classifying psoriasis were 0.872 and 0.838, 1.000 and 0.605, respectively. Both the dermatologists and CNN achieved at least moderate consistency with the reference standard, and there was no significant difference in Kappa coefficients between them (P > 0.05).ConclusionsThe performance of CNN developed with relatively modest number of dermoscopic images of skin tumors and psoriasis for Chinese population is comparable with 164 dermatologists. These two models could be used for screening in patients suspected with skin tumors and psoriasis respectively in primary care hospital.

Deep Learning Classifiers for Automated Detection of Gonioscopic Angle Closure Based on Anterior Segment OCT Images

To develop and test deep learning classifiers that detect gonioscopic angle closure and primary angle closure disease (PACD) based on fully automated analysis of anterior segment OCT (AS-OCT) images. Subjects were recruited as part of the Chinese-American Eye Study (CHES), a population-based study of Chinese Americans in Los Angeles, California, USA. Each subject underwent a complete ocular examination including gonioscopy and AS-OCT imaging in each quadrant of the anterior chamber angle (ACA). Deep learning methods were used to develop 3 competing multi-class convolutional neural network (CNN) classifiers for modified Shaffer grades 0, 1, 2, 3, and 4. Binary probabilities for closed (grades 0 and 1) and open (grades 2, 3, and 4) angles were calculated by summing over the corresponding grades. Classifier performance was evaluated by 5-fold cross-validation and on an independent test dataset. Outcome measures included area under the receiver operating characteristic curve (AUC) for detecting gonioscopic angle closure and PACD, defined as either 2 or 3 quadrants of gonioscopic angle closure per eye. A total of 4036 AS-OCT images with corresponding gonioscopy grades (1943 open, 2093 closed) were obtained from 791 CHES subjects. Three competing CNN classifiers were developed with a cross-validation dataset of 3396 images (1632 open, 1764 closed) from 664 subjects. The remaining 640 images (311 open, 329 closed) from 127 subjects were segregated into a test dataset. The best-performing classifier was developed by applying transfer learning to the ResNet-18 architecture. For detecting gonioscopic angle closure, this classifier achieved an AUC of 0.933 (95% confidence interval, 0.925-0.941) on the cross-validation dataset and 0.928 on the test dataset. For detecting PACD based on 2- and 3-quadrant definitions, the ResNet-18 classifier achieved AUCs of 0.964 and 0.952, respectively, on the test dataset. Deep learning classifiers effectively detect gonioscopic angle closure and PACD based on automated analysis of AS-OCT images. These methods could be used to automate clinical evaluations of the ACA and improve access to eye care in high-risk populations.

Joint Prostate Cancer Detection and Gleason Score Prediction in mp-MRI via FocalNet.

Multi-parametric MRI (mp-MRI) is considered the best non-invasive imaging modality for diagnosing prostate cancer (PCa). However, mp-MRI for PCa diagnosis is currently limited by the qualitative or semi-quantitative interpretation criteria, leading to inter-reader variability and a suboptimal ability to assess lesion aggressiveness. Convolutional neural networks (CNNs) are a powerful method to automatically learn the discriminative features for various tasks, including cancer detection. We propose a novel multi-class CNN, FocalNet, to jointly detect PCa lesions and predict their aggressiveness using Gleason score (GS). FocalNet characterizes lesion aggressiveness and fully utilizes distinctive knowledge from mp-MRI. We collected a prostate mp-MRI dataset from 417 patients who underwent 3T mp-MRI exams prior to robotic-assisted laparoscopic prostatectomy. FocalNet was trained and evaluated in this large study cohort with fivefold cross validation. In the free-response receiver operating characteristics (FROC) analysis for lesion detection, FocalNet achieved 89.7% and 87.9% sensitivity for index lesions and clinically significant lesions at one false positive per patient, respectively. For the GS classification, evaluated by the receiver operating characteristics (ROC) analysis, FocalNet received the area under the curve of 0.81 and 0.79 for the classifications of clinically significant PCa (GS ≥ 3 + 4) and PCa with GS ≥ 4 + 3, respectively. With the comparison to the prospective performance of radiologists using the current diagnostic guideline, FocalNet demonstrated comparable detection sensitivity for index lesions and clinically significant lesions, only 3.4% and 1.5% lower than highly experienced radiologists without statistical significance.

Two-Stage Convolutional Neural Network for Ship and Spill Detection Using SLAR Images

This paper presents a system for the detection of ships and oil spills using side-looking airborne radar (SLAR) images. The proposed method employs a two-stage architecture composed of three pairs of convolutional neural networks (CNNs). Each pair of networks is trained to recognize a single class (ship, oil spill, and coast) by following two steps: a first network performs a coarse detection, and then, a second specialized CNN obtains the precise localization of the pixels belonging to each class. After classification, a postprocessing stage is performed by applying a morphological opening filter in order to eliminate small look-alikes, and removing those oil spills and ships that are surrounded by a minimum amount of coast. Data augmentation is performed to increase the number of samples, owing to the difficulty involved in obtaining a sufficient number of correctly labeled SLAR images. The proposed method is evaluated and compared to a single multiclass CNN architecture and to previous state-of-the-art methods using accuracy, precision, recall, F-measure, and intersection over union. The results show that the proposed method is efficient and competitive, and outperforms the approaches previously used for this task.