Identifikasi Penyakit Pneumonia COVID-19 Berbasis Citra Computed Tomography (CT) Menggunakan Deep Transfer Learning

Abstract

Dunia sedang dilanda pandemi COVID-19 yang menyebabkan pneumonia hingga kematian pada penderitanya. Karena tingkat penyebaran virus SARS-CoV 2 yang tinggi, maka pemberian tes harus dilakukan secara cepat dan masif agar penderita dapat diisolasi sesegera mungkin. Pendekatan standar yang digunakan untuk mendeteksi COVID-19 saat ini adalah RT-PCR. Namun, terdapat beberapa kekurangan seperti sensitivitasnya hanya sebesar 60-70% dan biayanya relatif mahal. Penelitian ini bertujuan untuk membangun metode pendeteksian alternatif COVID-19 berbasis citra Computed Tomography menggunakan teknik deep transfer learning. Transfer learning dilakukan menggunakan pre-trained model DenseNet-201. Pembuatan model terdiri dari beberapa tahapan yaitu praproses data, pembagian data, augmentasi data, hyperparameter tuning, pembangunan model klasifikasi deep transfer learning, dan evaluasi. Model yang telah dibangun memiliki performa yang cukup baik dalam mengidentifikasi paru-paru dengan pneumonia COVID-19 dan paru-paru sehat dengan akurasi sebesar 93,41%, presisi sebesar 94,19%, sensitivitas (recall) sebesar 93,10%, dan spesifisitas sebesar 93,75%.

The world is being stricken by the COVID-19 pandemic, which can cause pneumonia, even the worst case is death. Due to the high level of spread of the SARS-CoV-2 virus, rapid and massive tests must be done. The infected people need to be isolated as soon as possible. The gold standard used to detect COVID-19 is RT-PCR. However, there are some drawbacks, this method only has 60-70% sensitivity, and has a relatively high cost. This study aims to develop an alternative detection method for COVID-19 based on Computed Tomography images using deep transfer learning techniques. Transfer learning is carried out using the pre-trained DenseNet-201 model. The process consists of several stages; data preprocessing, data splitting, data augmentation, hyperparameter tuning, developing a deep transfer learning classification model, and evaluation. The model that has been built has a fairly good performance in identifying lungs with pneumonia COVID-19 and healthy lungs with accuracy 93,41%, precision 94,19%, sensitivity (recall) 93,10%, and specificity of 93,75%.