Analisis Klasifikasi Citra Histopatologi Kanker Paru-paru Menggunakan Arsitektur DenseNet121

Abstract

Secara global, diperkirakan terdapat 2,4 juta kasus baru kanker paru-paru dan sekitar 1,8 juta kematian pada tahun 2022, mewakili 18,7% dari seluruh kematian akibat kanker. Kanker paru merupakan penyebab kematian utama, yaitu 14,1% (30,843 kematian) dari seluruh kematian akibat kanker di Indonesia. Penelitian ini bertujuan untuk membuat model convolutional neural network dengan arsitektur DenseNet121 dalam memprediksi status kanker paru-paru pada citra histopatologi. Data yang digunakan dalam penelitian ini bersumber dari dataset LungHist700 yang terdiri atas 691 citra histopatologi jaringan kanker paru paru dan terbagi menjadi tiga kelompok, yaitu normal, adenocarcinoma, dan squamous cell carcinoma. Arsitektur DenseNet121 digunakan untuk melatih model karena kemampuannya dalam mengurangi kebutuhan sumber daya komputasi yang tinggi. Pada penelitian ini, terdapat delapan kombinasi model yang dihasilkan dari kombinasi praproses contrast limited adaptive histogram equalization (CLAHE) dan tanpa CLAHE, serta hyperparameter. Model terbaik dihasilkan dari model dengan praproses CLAHE, learning rate 1e-5, batch size 16, yang mencapai akurasi 89,93%., precision 90,30%, sensitivitas 89,93%, F1-Score 89,97%, dan balanced accuracy 91,53%.

Globally, it is estimated that there were 2.4 million new cases of lung cancer and approximately 1.8 million deaths in 2022, representing 18.7% of all cancer related deaths. Lung cancer is the leading cause of cancer mortality in Indonesia, accounting for 14.1% (30,843 deaths) of all cancer deaths. This study aims to develop a convolutional neural network model with the DenseNet121 architecture to predict lung cancer status from histopathological images. The data used in this study comes from the LungHist700 dataset, which consists of 691 histopathological images of lung cancer tissue divided into three classes, normal, adenocarcinoma, and squamous cell carcinoma. The DenseNet121 architecture was chosen to train the model due to its ability to reduce the need for high computational resources. In this study, eight model combinations were created from the use of preprocessing with contrast limited adaptive histogram equalization (CLAHE) and without CLAHE, along with different hyperparameters. The best model was obtained from the combination using CLAHE preprocessing, a learning rate of 1e-5, and a batch size of 16, achieving an accuracy of 89.93%, precision of 90.30%, sensitivity of 89.93%, F1-Score of 89.97%, and balanced accuracy of 91.53%.