Penanganan Ketidakseimbangan Data Menggunakan Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) untuk Klasifikasi Citra Chest X-Ray Penyakit Pneumonia

Nareswari, Putri Hanna (2026) Penanganan Ketidakseimbangan Data Menggunakan Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) untuk Klasifikasi Citra Chest X-Ray Penyakit Pneumonia. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5003221082-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Pneumonia merupakan salah satu masalah kesehatan global utama yang memerlukan deteksi dini untuk menurunkan angka mortalitas. Citra Chest X-Ray (CXR) digunakan sebagai modalitas pencitraan untuk mendukung diagnosis pneumonia, tetapi interpretasi manual citra CXR berpotensi menyebabkan keterlambatan diagnosis. Convolutional Neural Network (CNN) berbasis transfer learning telah digunakan untuk klasifikasi pneumonia secara otomatis, tetapi performanya menurun pada kondisi data tidak seimbang antar kelas. Penelitian ini bertujuan menganalisis efektivitas penerapan Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) dalam mengatasi ketidakseimbangan data citra CXR serta mengevaluasi peningkatan kinerja klasifikasi multikelas pneumonia secara konsisten pada tiga arsitektur CNN, yaitu ResNet-50, DenseNet-201, dan Xception. Data yang digunakan dalam penelitian ini merupakan data sekunder berupa citra CXR dari Mendeley Data yang terdiri atas kelas normal, pneumonia bakterial, dan pneumonia viral. CWGAN-GP digunakan untuk membangkitkan citra sintetis pada kelas minoritas hingga distribusi data latih seimbang, kemudian dimanfaatkan untuk pelatihan ulang ketiga arsitektur. Hasil penelitian menunjukkan bahwa CWGAN-GP mampu menghasilkan citra sintetis yang menyerupai citra asli dengan nilai Fréchet Inception Distance (FID) keseluruhan terbaik sebesar 151,65. Proses pelatihan berlangsung stabil tanpa indikasi mode collapse maupun critic collapse. Penerapan CWGAN-GP meningkatkan kinerja klasifikasi secara konsisten pada ketiga arsitektur CNN. Peningkatan paling signifikan terjadi pada kelas pneumonia viral sebagai kelas minoritas, yang ditunjukkan oleh peningkatan nilai recall serta penurunan jumlah kesalahan klasifikasi sebagai normal (false negative) pada seluruh arsitektur. Model CWGAN-GP+DenseNet-201 menunjukkan kinerja klasifikasi terbaik dengan accuracy sebesar 0,93, precision sebesar 0,91, recall sebesar 0,92, F1-score sebesar 0,92, dan macro average AUC sebesar 0,98, serta hanya dua citra pneumonia viral yang salah diklasifikasikan sebagai normal. Penerapan CWGAN-GP terbukti efektif dalam mengatasi ketidakseimbangan data serta memberikan peningkatan kinerja klasifikasi multikelas pneumonia secara konsisten pada ketiga arsitektur CNN yang dievaluasi, sehingga berpotensi mendukung deteksi dini pneumonia berbasis citra CXR.
======================================================================================================================================
Pneumonia is one of the major global health problems that requires early detection to reduce mortality. Chest X-Ray (CXR) imaging is widely used as an imaging modality to support pneumonia diagnosis; however, manual interpretation of CXR images may lead to delayed diagnosis. Transfer learning-based Convolutional Neural Networks (CNNs) have been widely employed for automated pneumonia classification, but their performance tends to decline when dealing with class-imbalanced datasets. This study aims to analyze the effectiveness of the Conditional Wasserstein Generative Adversarial Network with Gradient Penalty (CWGAN-GP) in addressing class imbalance in CXR images and to evaluate its effectiveness in consistently improving multiclass pneumonia classification performance across three CNN architectures, namely ResNet-50, DenseNet-201, and Xception. The study used a secondary dataset consisting of CXR images obtained from Mendeley Data, comprising three classes: normal, bacterial pneumonia, and viral pneumonia. CWGAN-GP was employed to generate synthetic images for the minority class until a balanced training dataset was achieved, after which the augmented dataset was used to retrain the three CNN architectures. The results demonstrate that CWGAN-GP successfully generated synthetic images that closely resemble real CXR images, achieving the best overall Fréchet Inception Distance (FID) score of 151.65. The training process remained stable without indications of mode collapse or critic collapse. The application of CWGAN-GP consistently improved the classification performance across all three CNN architectures. The most significant improvement was observed in the minority viral pneumonia class, as indicated by an increase in recall and a reduction in false negative predictions across all architectures. The CWGAN-GP+DenseNet-201 model achieved the best classification performance, with an accuracy of 0.93, precision of 0.91, recall of 0.92, an F1-score of 0.92, and a macro-average AUC of 0.98, while only two viral pneumonia images were misclassified as normal. The findings demonstrate that CWGAN-GP is effective in addressing class imbalance and consistently improving multiclass pneumonia classification performance across the three evaluated CNN architectures, thereby demonstrating its potential to support early pneumonia detection using CXR images.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Chest X-Ray, CNN, CWGAN-GP, Ketidakseimbangan Data, Pneumonia, Data Imbalance, Pneumonia
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Science and Data Analytics (SCIENTICS) > Statistics > 49201-(S1) Undergraduate Thesis
Depositing User:	Putri Hanna Nareswari
Date Deposited:	29 Jul 2026 07:00
Last Modified:	29 Jul 2026 07:00
URI:	http://repository.its.ac.id/id/eprint/139816

Actions (login required)

View Item