Publicação
Local analysis strategies for exudate detection in fundus images
| Resumo: | Diabetic Retinopathy (DR) is a common complication of diabetes, which is among the major causes of vision loss in the world. An early detection of the disease is the key to avoid the patient’s blindness. However, at the initial phase of the disease, the vision impairment is not easily percieved by the patient. Therefore, regular follow-up exams are recommended in order to detect anomalous patterns in the patient’s retina. Exudates are one of the most prevalent signs during the early stage of DR and, therefore, its early detection is vital to prevent the patient’s blindness. However, the manual detection of exudates by experts is laborious and time-consuming. Thus, automated screening techniques for exudate detection have great significance in saving cost, time and labor, allowing the ophthalmologists to make the treatment decision timely. In this sense, one of the main objectives of this thesis is to develop and compare different strategies to locally extract information of fundus images for detecting exudates. Several methods related to the automatic detection of exudates have been proposed in the literature however, these methods focus their efforts in the segmentation of exudates or require the extraction of features from a lesion candidate map. On the other hand, in the methodologies proposed in this thesis, the characterization of healthy and damaged retinal areas is performed by applying image descriptors in a local way, avoiding the segmentation step and the generation of candidate maps. A system based on local feature extraction and Support Vector Machine classification is used to develop and compare different strategies for automated detection of exudates. The main novelty of this work is allowing the detection of exudates using non-regular regions to perform the local feature extraction. To accomplish this objective, different methods for generating superpixels are applied to the fundus images of E-OPHTA database and texture and morphological features are extracted for each of the resulting regions. Finally, each region is classified according to healthy and pathological classes, during the classification stage. The strategies proposed in order to generate superpixels rely on applying the marker-controlled watershed transformation to a spatially regularized gradient. From these strategies, two different types of superpixels are created: c-Waterpixels and m-Waterpixels. In the end, an elaborated comparison between the proposed methods for generating m and c-waterpixels and the state-of-the-art method for generating SLIC superpixels is performed. Additionally, a system based on Convolutional Neural Networks (CNN) is explored to discriminate between healthy and pathological regions in fundus images. Transfer learning is applied to fine-tune some of the most important state-of-the-art CNN architectures. Exudates usually represent less than one percent of the total number of pixels that compose the retinal image. This is the reason why, in both the systems presented in this thesis, the fundus images are divided in superpixels and the classification is performed for each of the regions. Lastly, an exhaustive comparison between the two created systems to automatically detect exudates is performed. In other words, the classification results obtained through the system involving CNNs are compared with the ones obtained by applying the approach based on feature extraction and subsequent classification using machine learning algorithms. |
|---|---|
| Autores principais: | Pereira, Joana Daniela da Silva |
| Assunto: | Engenharia e Tecnologia::Engenharia Médica |
| Ano: | 2017 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade do Minho |
| Idioma: | inglês |
| Origem: | RepositóriUM - Universidade do Minho |
| Resumo: | Diabetic Retinopathy (DR) is a common complication of diabetes, which is among the major causes of vision loss in the world. An early detection of the disease is the key to avoid the patient’s blindness. However, at the initial phase of the disease, the vision impairment is not easily percieved by the patient. Therefore, regular follow-up exams are recommended in order to detect anomalous patterns in the patient’s retina. Exudates are one of the most prevalent signs during the early stage of DR and, therefore, its early detection is vital to prevent the patient’s blindness. However, the manual detection of exudates by experts is laborious and time-consuming. Thus, automated screening techniques for exudate detection have great significance in saving cost, time and labor, allowing the ophthalmologists to make the treatment decision timely. In this sense, one of the main objectives of this thesis is to develop and compare different strategies to locally extract information of fundus images for detecting exudates. Several methods related to the automatic detection of exudates have been proposed in the literature however, these methods focus their efforts in the segmentation of exudates or require the extraction of features from a lesion candidate map. On the other hand, in the methodologies proposed in this thesis, the characterization of healthy and damaged retinal areas is performed by applying image descriptors in a local way, avoiding the segmentation step and the generation of candidate maps. A system based on local feature extraction and Support Vector Machine classification is used to develop and compare different strategies for automated detection of exudates. The main novelty of this work is allowing the detection of exudates using non-regular regions to perform the local feature extraction. To accomplish this objective, different methods for generating superpixels are applied to the fundus images of E-OPHTA database and texture and morphological features are extracted for each of the resulting regions. Finally, each region is classified according to healthy and pathological classes, during the classification stage. The strategies proposed in order to generate superpixels rely on applying the marker-controlled watershed transformation to a spatially regularized gradient. From these strategies, two different types of superpixels are created: c-Waterpixels and m-Waterpixels. In the end, an elaborated comparison between the proposed methods for generating m and c-waterpixels and the state-of-the-art method for generating SLIC superpixels is performed. Additionally, a system based on Convolutional Neural Networks (CNN) is explored to discriminate between healthy and pathological regions in fundus images. Transfer learning is applied to fine-tune some of the most important state-of-the-art CNN architectures. Exudates usually represent less than one percent of the total number of pixels that compose the retinal image. This is the reason why, in both the systems presented in this thesis, the fundus images are divided in superpixels and the classification is performed for each of the regions. Lastly, an exhaustive comparison between the two created systems to automatically detect exudates is performed. In other words, the classification results obtained through the system involving CNNs are compared with the ones obtained by applying the approach based on feature extraction and subsequent classification using machine learning algorithms. |
|---|