Deep neural networks have achieved great success in many domains. However, successful deployment of such systems is determined by proper manual selection of the neural architecture. This is a tedious and time-consuming process that requires expert knowledge. Different tasks need very different architectures to obtain satisfactory results. The group of methods called the neural architecture search (NAS) helps to find effective architecture in an automated manner. In this paper, we present the use of an architecture search framework to solve the medical task of malignant melanoma detection. Unlike many other methods tested on benchmark datasets, we tested it on practical problem, which differs greatly in terms of difficulty in distinguishing between classes, resolution of images, data balance within the classes, and the number of data available. In order to find a suitable network structure, the hill-climbing search strategy was employed along with network morphism operations to explore the search space. The network morphism operations allow for incremental increases in the network size with the use of the previously trained network. This kind of knowledge reusing allows significantly reducing the computational cost. The proposed approach produces structures that achieve similar results to those provided by manually designed structures, at the same time making use of almost 20 times fewer parameters. What is more, the search process lasts on average only 18h on single GPU.
The paper presents the results of the research on
neural architecture search (NAS) algorithm. We utilized the hill
climbing algorithm to search for well-performing structures of
deep convolutional neural network. Moreover, we used the
function preserving transformations which enabled the effective
operation of the algorithm in a short period of time. The network
obtained with the advantage of NAS was validated on skin lesion
classification problem. We compared the parameters and
performance of the automatically generated neural structure with
the architectures selected manually, reported by the authors in
previous papers. The obtained structure achieved comparable
results to hand-designed networks, but with much fewer
parameters then manually crafted architectures.
Malignant melanomas are the most deadly type of skin cancer but detected early have high chances for successful treatment. In the last twenty years, the interest of automated melanoma recognition detection and classification dynamically increased partially because of public datasets appearing with dermatoscopic images of skin lesions. Automated computer-aided skin cancer detection in dermatoscopic images is a very challenging task due to uneven datasets sizes, the huge intra-class variation with small interclass variation, and the existence of many artifacts in the image. One of the most recognized methods of melanoma diagnosis is the ABCD method. In the paper, we propose an extended version of this method and an intelligent decision support system based on neural networks that uses its results in a form of hand-crafted features. Automatic determination of the skin features used by the ABCD method is difficult due to the large diversity of images of various quality, the existence of hair, different markers and other obstacles. Therefore, it was necessary to apply advanced methods of preprocessing the images. The system is an ensemble of ten neural networks, working in parallel and one network using their results to generate a final decision. This system structure allowed us to increase the efficiency of the operation by several percentage points compared to a single neural network. The proposed system is trained on over 5000 and tested afterward on 200 skin moles. The presented system can be used as a decision support system for primary care physicians, as a system capable of self-examination of the skin with a dermatoscope and also as an important tool to improve biopsy decision making.
In recent years, deep learning and especially Deep Neural Networks (DNN) have obtained amazing performance on a variety of problems, in particular in classification or pattern recognition. Among many kinds of DNNs, the Convolutional Neural Networks (CNN) are most commonly used. However, due to their complexity, there are many problems related but not limited to optimizing network parameters, avoiding overfitting and ensuring good generalization abilities. Therefore, a number of methods have been proposed by the researchers to deal with these problems. In this paper, we present the results of applying different, recently developed methods to improve deep neural network training and operating. We decided to focus on the most popular CNN structures, namely on VGG based neural networks: VGG16, VGG11 and proposed by us VGG8. The tests were conducted on a real and very important problem of skin cancer detection. A publicly available dataset of skin lesions was used as a benchmark. We analyzed the influence of applying: dropout, batch normalization, model ensembling, and transfer learning. Moreover, the influence of the type of activation function was checked. In order to increase the objectivity of the results, each of the tested models was trained 6 times and their results were averaged. In addition, in order to mitigate the impact of the selection of learning, test and validation sets, k-fold validation was applied.
The diabetic retinopathy is a disease caused by long-standing diabetes. Lack of effective treatment can lead to vision impairment and even irreversible blindness. The disease can be diagnosed by examining digital color fundus photographs of retina. In this paper we propose deep learning approach to automated diabetic retinopathy screening. Deep convolutional neural networks (CNN) - the most popular kind of deep learning algorithms - enjoyed great success in the field of image analysis and recognition. Therefore, we leverage CNN networks to diagnose the diabetic retinopathy and its current stage, based on analysis of the photographs of retina. The utilized models were trained using dataset containing over 88000 retina photographs, labeled by specialist clinicians. To enhance the performance of the system, we proposed a special class coding technique that enabled to include the information about value of difference between predicted score and target score into the objective function being minimized during the neural networks training. To evaluate classification ability of employed models we used standard accuracy metrics and quadratic weighted Kappa score that is calculated between the predicted scores and scores provided in the dataset. The best tested model achieved an accuracy of about 82% in detecting the retinopathy and 51% in assessing its stage. Moreover, system obtained decent Kappa score equal 0.776. Achieved results showed that deep learning algorithms can be successfully employed to solve this very hard to analyze problem.
W referacie opisano problem wykrywania oraz klasyfikacji stanu retinopatii cukrzycowej ze zdjęć dna oka przy pomocy głębokich sieci neuronowych. Retinopatia cukrzycowa jest chorobą oczu często występującą u osób z cukrzycą. Nieleczona prowadzi do uszkodzenia wzroku, a nawet ślepoty. W pracy badawczej opracowano system wykrywania retinopatii cukrzycowej na podstawie zdjęć dna oka. Opracowana sieć neuronowa przypisuje stan choroby w 5 stopniowej skali – od braku choroby do najbardziej zaawansowanego stanu choroby. Zaproponowano specjalny system kodowania klas w celu uchwycenia wielkości różnicy pomiędzy rzeczywistymi a predykowanymi stanami choroby. Uzyskano wysokie wyniki klasyfikacji na zbiorze testowym. W celu oceny skuteczności działania systemu wykorzystano miary statystyczne takie jak ważona Kappa i dokładność.
Malignant melanomas are the most deadly type of skin cancers however detected early enough give a high chances for successful treatment. The last years saw the dynamic growth of interest of automatic computer-aided skin cancer diagnosis. Every month brings new research results on new approaches to this problem, new methods of preprocessing, new classifiers, new ideas to follow etc. In particular, the rapid development of dermatoscopy, image processing methods, as well as the ever-increasing computing power of computers caused that researchers are able to consider significantly more features of the analyzed lesion than has been done so far using methods recognized in a medical community such as ABCD or Menzies methods. From the other hand more features not always imply an improvement in terms of efficiency of the diagnosis and its transparency. Hence, in this paper we survey the kind of features taken into account by the researchers and then, selected the most efficient set of them. Proposed method jointly selects the optimal set of features representing the analyzed lesion together with the accompanying form of the neural classifier (number of neurons, activation functions). The evolutionary algorithms are used in order to carry out the optimization. Obtained results are even better than the ones obtained by the most efficient these days deep classifiers.
W referacie zaprezentowane zostaną wyniki badań nad rozpoznawaniem obiektów w różnych warunkach za pomocą głębokich sieci neuronowych. Przeanalizowano działanie dwóch struktur – ResNet50 oraz VGG19. Systemy rozpoznawania obrazu wytrenowano oraz przetestowano na reprezentatywnej, bazie zawierającej 25 tys. zdjęć psów oraz kotów, która znacznie upraszcza analizowanie działania systemów ze względu na łatwość interpretacji zdjęć przez człowieka. Zbadano wpływ pojawienia się nietypowych zdjęć na wynik klasyfikacji. Ponadto przeanalizowano zdjęcia niepoprawnie sklasyfikowane i porównano je z interpretacjami człowieka. Uzyskano bardzo wysokie wyniki klasyfikacji. Do oceny systemów użyto miar statystycznych takich jak: dokładność, czułość, swoistość, krzywe ROC
The paper presents utilization of one of the latest tool from the group of Machine learning techniques, namely Deep Convolutional Neural Networks (CNN), in process of decision making in selected medical problems. After the survey of the most successful applications of CNN in solving medical problems, the paper focuses on the very difficult problem of automatic analyses of the skin lesions. The authors propose the CNN structure and the way to cope with the insufficient number of learning data. The research was carried out and validated on the data base of over 10000 images. The efficiency of the proposed approach reaches 84%.
The paper presents the results of research on the use of Deep Neural Networks (DNN) for automatic classification of the skin lesions. The authors have focused on the most effective kind of DNNs for image processing, namely Convolutional Neural Networks (CNN). In particular, three kinds of CNN were analyzed: VGG19, Residual Networks (ResNet) and the hybrid of VGG19 CNN with the Support Vector Machine (SVM). The research was carried out with the use of database of over 10 000 images representing skin lesions: benign and malignant. Because of an uneven number of images representing different classes of lesions, the up-sampling of underrepresented class was applied. The comparison of the CNN structures with respect to the accuracy, sensitivity and specificity was performed using k-fold validation method.