KLASIFIKASI KANKER KULIT MENGGUNAKAN TEKNIK DEEP TRANSFER LEARNING BERDASARKAN CITRA DERMOSKOPI

Qorry Aina Fitroh, NIM.: 21206051016 (2023) KLASIFIKASI KANKER KULIT MENGGUNAKAN TEKNIK DEEP TRANSFER LEARNING BERDASARKAN CITRA DERMOSKOPI. Masters thesis, UIN SUNAN KALIJAGA YOGYAKARTA.

Preview	Text (KLASIFIKASI KANKER KULIT MENGGUNAKAN TEKNIK DEEP TRANSFER LEARNING BERDASARKAN CITRA DERMOSKOPI) 21206051016_BAB-I_IV-atau-V_DAFTAR-PUSTAKA.pdf - Published Version Download (3MB) | Preview
	Text (KLASIFIKASI KANKER KULIT MENGGUNAKAN TEKNIK DEEP TRANSFER LEARNING BERDASARKAN CITRA DERMOSKOPI) 21206051016_BAB-II_sampai_SEBELUM-BAB-TERAKHIR.pdf - Published Version Restricted to Registered users only Download (4MB) | Request a copy

Abstract

Skin cancer is common cancer in humans and can cause death. Handling early symptoms with the proper treatment can help patients to have a greater chance of life. The accuracy of early diagnosis in skin cancer patients plays an important role, so it needs a technique that can accurately classify skin cancer types. Various deep-learning techniques have been developed to perform skin cancer diagnosis, but still require optimization to improve detection accuracy. This study proposes two deep learning techniques to optimize the classification of dermoscopic images of skin cancer in two classes (benign and malignant) using pre-trained models VGG-16 and ResNet-50. The first technique is called fine-tuning which is done by adding two dense layers with ReLu activation function (512 and 256 neurons) and one dense layer with softmax activation function and two neurons. The fine-tuning technique is tested through four tuning hyperparameters - optimization, batch size, learning rate, and epoch. The second technique is called off-the-shelf feature extraction, which replaces the fully connected layer in the pre-trained CNN model with a Support Vector Machine (SVM) classifier. Furthermore, optimization is carried out on SVM with two methods, grid search and tree-based pipeline optimization tool (TPOT). This research resulted in the best accuracy using the ResNet-50 pre-trained model which is 94% for the first technique and 91.5% accuracy obtained with the second technique using the VGG-16 pre-trained model.

Share this knowledge with your friends :

Actions (login required)

View Item