| Summary: | High sensitivity and accuracy result in detection and classification improved the chances of survival for lung cancer patients significantly. To accomplish this goal, Computer-Aided Detection (CAD) system using the CNN deep learning method has been developed. In this study, we propose a modified ResNet50 architecture and transfer learning to classify lung cancer images into four classes. The modification of ResNet50 was to replace the last layer, which was a global average pooling layer with two layers, namely a flatten and dense layer. In addition, we also added a zero-padding layer to the feature extraction process. We obtained results from the modified ResNet50 architecture are 93.33 accuracy, 92.75 sensitivity, 93.75 precision, 93.25 F1-score, and 0.559 of AUC. In this study, we also compared the modified ResNet50 with two other deep learning architectures: EfficientNetB1 and AlexNet. We used Kaggle public datasets, which contain 899 for training and validation, and 97 for testing, and an image of a CT scan that had already been labeled and classified. From our work, the modified ResNet50 architecture is the best in classifying lung cancer images into four classes (adenocarcinoma, large carcinoma, normal and squamous carcinoma) compared to the other two architectures. © 2021 IEEE.
|
|---|