Analisis Klasifikasi Citra Histopatologi Kanker Paru-Paru Menggunakan Arsitektur ResNet50
Abstract
Kanker paru merupakan salah satu penyebab utama kematian akibat kanker di dunia. Identifikasi tipe histologis kanker paru melalui citra histopatologi berperan penting dalam penentuan diagnosis dan penanganan pasien, tetapi interpretasinya membutuhkan keahlian ahli patologi dan dapat dipengaruhi variasi antar-pengamat. Penelitian ini bertujuan mengembangkan model klasifikasi citra histopatologi kanker paru-paru menggunakan arsitektur ResNet50 berbasis transfer learning. Dataset LungHist700 terdiri atas 691 citra dalam tiga kelas, yaitu adenocarcinoma (ACA), normal, dan squamous cell carcinoma (SCC), yang dibagi secara stratified menjadi 80% data training, 10% data validation, dan 10% data testing. Praproses meliputi resizing, normalisasi, serta augmentasi berupa random rotation, resized crop, horizontal flip, vertical flip, adjust sharpness, dan color jitter. Training dilakukan dalam dua tahap, yaitu pembekuan feature extractor dan fine-tuning seluruh layer, dengan 27 kombinasi batch size, learning rate, dan dropout. Konfigurasi terbaik, yaitu batch size 24, learning rate 1 × 10-4, dan dropout 0,4, menghasilkan accuracy 95,71%, precision 95,75%, sensitivity 95,71%, specificity 97,83%, dan F1-score 95,70% serta mengklasifikasikan 67 dari 70 citra testing secara benar. Hasil ini menunjukkan bahwa ResNet50 berbasis transfer learning mampu mengklasifikasikan tiga kelas citra histopatologi kanker paru-paru dengan baik. Lung cancer is one of the leading causes of cancer-related deaths worldwide. Identifying the histological type of lung cancer from histopathological images is important for diagnosis and patient management, but interpretation requires pathologists’ expertise and may be affected by inter-observer variability. This study aimed to develop a lung cancer histopathological image classification model using a transfer learning-based ResNet50 architecture. The LungHist700 dataset consists of 691 images in three classes: adenocarcinoma (ACA), normal, and squamous cell carcinoma (SCC), divided using a stratified split into 80% training, 10% validation, and 10% testing data. Preprocessing included resizing, normalization, and augmentation using random rotation, resized crop, horizontal flip, vertical flip, adjust sharpness, and color jitter. Training was conducted in two stages, freezing the feature extractor and fine-tuning all layers, across 27 combinations of batch size, learning rate, and dropout. The best configuration, with a batch size of 24, a learning rate of 1 × 10-4, and a dropout rate of 0.4, achieved an accuracy of 95.71%, precision of 95.75%, sensitivity of 95.71%, specificity of 97.83%, and F1-score of 95.70%, correctly classifying 67 of 70 testing images. These results indicate that transfer learning-based ResNet50 effectively classifies three classes of lung cancer histopathological images.
Collections
- UF - Computer Science [197]

