FIRMANSYAH, IHSAN ALIF (2026) ANALISIS KOMPARASI ARSITEKTUR CNN: RESNET-50, DENSENET-121, DAN EFFICIENTNET-B0 UNTUK KLASIFIKASI KANKER KULIT. S1 thesis, Universitas Mercu Buana Jakarta.
|
Text (HAL COVER)
Cover.pdf Download (1MB) | Preview |
|
![[img]](https://repository.mercubuana.ac.id/style/images/fileicons/text.png) |
Text (BAB I)
Bab 1.pdf Restricted to Registered users only Download (35kB) |
|
|
Text (BAB II)
Bab 2.pdf Restricted to Registered users only Download (283kB) |
|
|
Text (BAB III)
Bab 3.pdf Restricted to Registered users only Download (628kB) |
|
|
Text (BAB IV)
Bab 4.pdf Restricted to Registered users only Download (664kB) |
|
|
Text (BAB V)
Bab 5.pdf Restricted to Registered users only Download (26kB) |
|
|
Text (DAFTAR PUSTAKA)
Daftar Pustaka.pdf Restricted to Registered users only Download (172kB) |
|
|
Text (LAMPIRAN)
Lampiran.pdf Restricted to Registered users only Download (348kB) |
Abstract
Skin cancer is one of the most prevalent types of cancer, with the number of cases continuing to increase over time. This condition highlights the importance of early detection to improve the chances of successful treatment. However, conventional diagnostic methods still rely on clinical examinations and biopsies, which require considerable time, high costs, and the expertise of medical professionals. This study aims to compare the performance of three Convolutional Neural Network (CNN) architectures, namely ResNet-50, DenseNet-121, and EfficientNet-B0, for classifying skin cancer into two categories: benign and malignant. The dataset used in this study was ISIC 2019, which originally consisted of eight diagnostic classes and was subsequently grouped into a binary classification. The preprocessing stage included image enhancement using Contrast Limited Adaptive Histogram Equalization (CLAHE) and resizing all images to 224 × 224 pixels. To address the class imbalance problem, offline data augmentation was applied to balance the training data distribution, followed by online data augmentation during the model training process. All models were developed using a transfer learning approach with ImageNet pretrained weights and trained using the AdamW optimizer with the Binary Cross-Entropy loss function combined with label smoothing. Model performance was evaluated using Accuracy, Balanced Accuracy, Precision, Recall, F1-Score, ROC-AUC, confusion matrix, as well as threshold tuning and hyperparameter tuning. The experimental results indicate that ResNet-50 achieved the best performance, with an accuracy of 87.22% and a ROC-AUC of 0.939. EfficientNet-B0 ranked second, achieving an accuracy of 86.60% and a ROC-AUC of 0.939, while DenseNet-121 obtained an accuracy of 84.45% and a ROC-AUC of 0.924. These findings demonstrate that ResNet-50 is the most effective architecture for skin cancer classification in this study and has the potential to serve as a reference for the development of deep learning-based medical image classification systems. Keywords: CNN, CLAHE, ResNet-50, DenseNet-121, EfficientNet-B0 Kanker kulit merupakan salah satu penyakit kanker yang jumlah penderitanya terus mengalami peningkatan dari waktu ke waktu. Kondisi ini menuntut adanya metode deteksi dini yang mampu meningkatkan peluang keberhasilan pengobatan. Meskipun demikian, proses diagnosis yang umum dilakukan masih bergantung pada pemeriksaan klinis dan biopsi, yang membutuhkan waktu relatif lama, biaya yang tinggi, serta keterlibatan tenaga medis yang berpengalaman. Penelitian ini bertujuan membandingkan kinerja tiga arsitektur Convolutional Neural Network (CNN), yaitu ResNet-50, DenseNet-121, dan EfficientNet-B0, dalam melakukan klasifikasi kanker kulit menjadi dua kategori, yaitu jinak dan ganas. Dataset yang digunakan berasal dari ISIC 2019 yang semula terdiri atas delapan kelas diagnosis, kemudian dikelompokkan menjadi klasifikasi biner. Tahap prapemrosesan meliputi peningkatan kualitas citra menggunakan Contrast Limited Adaptive Histogram Equalization (CLAHE) serta perubahan ukuran citra menjadi 224 × 224 piksel. Untuk mengatasi ketidakseimbangan jumlah data, dilakukan augmentasi offline hingga distribusi kelas menjadi seimbang, kemudian diterapkan augmentasi online selama proses pelatihan model. Seluruh model dikembangkan dengan pendekatan transfer learning menggunakan bobot awal ImageNet, kemudian dilatih menggunakan optimizer AdamW dan fungsi kehilangan Binary Cross Entropy yang dipadukan dengan label smoothing. Evaluasi performa dilakukan menggunakan metrik Accuracy, Balanced Accuracy, Precision, Recall, F1-Score, ROC-AUC, confusion matrix, serta melalui proses threshold tuning dan hyperparameter tuning. Hasil pengujian menunjukkan bahwa ResNet-50 memberikan performa terbaik dengan nilai accuracy sebesar 87,22% dan ROC-AUC sebesar 0,939. EfficientNet-B0 menempati urutan kedua dengan accuracy 86,60% dan ROC-AUC 0,939, sedangkan DenseNet-121 memperoleh accuracy 84,45% dan ROC-AUC 0,924. Berdasarkan hasil tersebut, dapat disimpulkan bahwa ResNet-50 merupakan arsitektur yang paling efektif untuk klasifikasi kanker kulit pada penelitian ini dan berpotensi dijadikan acuan dalam pengembangan sistem klasifikasi citra medis berbasis deep learning. Kata Kunci : CNN, CLAHE, ResNet-50, DenseNet-121, EfficientNet-B0
Actions (login required)
 |
View Item |
