Klasifikasi Diabetic Retinopathy dari Citra Fundus Berbasis Convolutional Neural Network

Alvionita, Vina (2020) Klasifikasi Diabetic Retinopathy dari Citra Fundus Berbasis Convolutional Neural Network. Undergraduate thesis, Institut Teknologi Sepuluh Nopember.

[img] Text
07311640000016-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only

Download (3MB) | Request a copy

Abstract

Diabetic retinopathy (DR) merupakan komplikasi diabetes yang paling sering ditemukan. DR terbagi menjadi beberapa tahapan yaitu mild NPDR, moderate NPDR, severe NPDR, dan PDR. Menilai tingkat keparahan ini bertujuan untuk menentukan treatment yang sesuai. Beberapa penelitian mengenai diagnosis DR secara otomatis berbasis Computer-aided Diagnosis (CADx) telah dilakukan. Metode tersebut menggunakan berbagai modul ekstraksi fitur dan dimasukkan dalam classifier tertentu. Metode berbasis fitur ekstraksi tersebut membutuhkan langkah yang panjang. Dilain sisi, deep neural networks telah berhasil diterapkan di berbagai bidang seperti klasifikasi citra dan menunjukkan kinerja yang baik. Untuk itu pada tugas akhir ini diajukan sebuah sistem klasifikasi DR secara otomatis melalui citra fundus berbasis Convolutional Neural Networks (CNN). Metode yang diusulkan terdiri dari beberapa tahapan yaitu preprocessing, data augmentation, dan classification. Dimana proses klasifikasi menggunakan Convolutional Neural Network. Pada tahap preprocessing terdiri dari 4 proses, yaitu normalisasi bentuk, resize, normalisasi warna, dan contrast enhancement. Dataset yang digunakan pada penelitian diperoleh dari Asia-Pasific Tele-Opthalmology Society (APTOS). Pembelajaran CNN dilakukan pada tiga kondisi yang berbeda. Kondisi pertama menggunakan dataset dengan jumlah data tidak seimbang. Kondisi kedua menggunakan dataset seimbang dengan melakukan undersampling pada jumlah kelas terendah. Kondisi ketiga menggunakan jumlah data yang seimbang dengan melakukan oversampling menyesuaikan jumlah kelas terbanyak. Dari ketiga pengujian, dengan memperhatikan seluruh performa evaluasi diperoleh hasil terbaik pada kondisi ketiga, dengan nilai akurasi 73,64%, presisi 59,01%, sensitivitas 60,69%, dan spesifisitas 93,49%. Kemudian dari pengujian menggunakan metode 5-fold cross validation, model 5 mencapai hasil terbaik dengan akurasi 79,36%, presisi 71,78%, sensitivitas 52,26%, dan spesifisitas 94,35%. ================================================================================================================== Diabetic retinopathy is the most common complication that occurs from a diabetes patient. DR is divided into several stages, namely mild NPDR, moderate NPDR, severe NPDR, and PDR. Assessing the severity of this aims to determine the appropriate treatment. Several studies on the diagnosis of DR automatically based on Computer-aided Diagnosis (CADx) have been conducted. The method uses various feature extraction modules and was entered in a particular classifier. The feature-based extraction method requires a long step. On the other hand, deep neural networks have been successfully applied in various fields such as image classification and showing good performance. Despite the progress of the study, there is still space for the development of this research. For this reason, this final project proposes an automatic classification system for DR through a fundus image based on Convolutional Neural Networks (CNN). The proposed method consists of several stages, namely preprocessing, data augmentation, and classification. The classification process uses the Convolutional Neural Network. In the preprocessing stage consists of 4 steps, namely shape normalization, resizing, color normalization, and contrast enhancement. The proposed system uses a dataset from the Asia-Pacific Tele-Ophthalmology Society (APTOS). CNN learning is carried out under three different conditions. The first condition uses an unbalanced dataset. The second condition uses a balanced dataset by undersampling the lowest number of classes. The third condition uses a balanced amount of data by oversampling to adjust the highest number of classes. From the three tests, the best results were obtained in the third condition, with an accuracy of 73.64%, precision 59.01%, sensitivity 60.69%, and specificity 93.49%. Then, from the tests use the 5 fold cross-validation method, model 5 achieved the best results with an accuracy of 79.36%, precision 71.78%, sensitivity 52.26%, and specificity 94.35%.

Item Type: Thesis (Undergraduate)
Uncontrolled Keywords: convolutional neural network, diabetic retinopathy, fundus image, image classification, convolutional neural network, diabetic retinopathy, fundus image, image classification
Subjects: T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques
Divisions: Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Biomedical Engineering > 11410-(S1) Undergraduate Thesis
Depositing User: Alvionita Vina
Date Deposited: 25 Aug 2020 02:49
Last Modified: 25 Aug 2020 02:49
URI: http://repository.its.ac.id/id/eprint/80244

Actions (login required)

View Item View Item