Deep Learning Image Classification Rontgen Dada pada Kasus Covid-19 Menggunakan Algoritma Convolutional Neural Network

Abstrak

Penelitian ini mengusulkan penggunaan Convolutional Neural Network (CNN) dengan arsitektur VGGNet-19 dan ResNet-50 untuk diagnosis COVID-19 melalui analisis citra rontgen dada. Modifikasi dilakukan dengan membandingkan nilai regularisasi dropout 50% dan 80% untuk kedua arsitektur dan mengubah jumlah lapisan klasfikasi menjadi 4 kelas. Selanjutnya, kinerja model dibandingkan berdasarkan ukuran dataset. Dataset terdiri dari 21165 citra, dengan pembagian 10% sebagai data uji dan 90% data dibagi menjadi data latih (80%) dan data validasi (20%). Kinerja model dievaluasi menggunakan metode validasi silang berulang 5 kali lipat. Proses pelatihan menggunakan learning rate 0.0001, optimasi stochastic gradient descent (SGD), dan sepuluh iterasi. Hasil penelitian menunjukkan bahwa penambahan lapisan dropout dengan peluang 50% untuk kedua arsitektur secara efektif mengatasi overfitting dan meningkatkan performa model. Ditemukan bahwa kinerja yang lebih baik dicapai pada ukuran kumpulan data lebih besar dan memberikan peningkatan signifikan pada kinerja model. Hasil klasifikasi menunjukkan arsitektur ResNet-50 mencapai akurasi rata-rata 94.4%, recall rata-rata 94.1%, presisi rata-rata 95.5%, spesifisitas rata-rata 97% dan F1-score rata-rata 94.8%. Sedangkan arsitektur VGGNet-19 mencapai akurasi rata-rata 91%, recall rata-rata 89%, presisi rata-rata 95.0%, spesifisitas rata-rata 96.8% dan F1-score rata-rata 92.7%. Pemanfaatan model ini dapat membantu mengidentifikasi penyebab kematian pasien dan memberikan informasi yang berharga bagi pengambilan keputusan medis dan epidemiologi.

Abstract

This research proposes using a Convolutional Neural Network (CNN) with VGGNet-19 and ResNet-50 architectures for COVID-19 diagnosis through chest X-ray image analysis. Modifications were made by comparing the dropout regularization values of 50% and 80% for both architectures and altering the number of classification layers to 4 classes. Furthermore, the model's performance was compared based on dataset size. The dataset comprised 21,165 images, with a division of 10% for testing and 90% divided into training data (80%) and validation data (20%). The model's performance was evaluated using the 5-fold repeat cross-validation method. The training process employed a learning rate of 0.0001, stochastic gradient descent (SGD) optimization, and ten iterations. The study's results indicate that adding dropout layers with a 50% probability for both architectures effectively addressed overfitting and improved the model's performance. It was found that better performance was achieved with larger dataset sizes. The classification results indicate the ResNet-50 architecture achieved an average accuracy of 94.4%, average recall of 94.1%, average precision of 95.5%, average specificity of 97%, and average F1-score of 94.8%. Meanwhile, the VGGNet-19 architecture achieved an average accuracy of 91%, an average recall of 89%, average precision of 95.0%, average specificity of 96.8%, and an average F1-score of 92.7%. Utilizing these models can assist in identifying the causes of patient mortality and offer valuable information for medical and epidemiological decision-making.