LUNG CANCER CLASSIFICATION FROM CT-SCAN IMAGES USING AN EFFICIENT CNN WITH TRANSFER LEARNING

Abstract

Lung cancer remains one of the leading causes of mortality globally, and early diagnosis plays a critical role in improving patient treatment and reducing mortality rates. However, the interpretation of medical images, such as Computed Tomography (CT) scans, is highly complex and demands significant clinical expertise. This research focuses on the fine-tuning an efficient Convolutional Neural Network (CNN) based approach, with particular emphasis on the EfficientNet-B0, for detecting lung cancer from CT scans into three classes: Normal, Malignant and Benign. The training data contains 1,097 CT scan slices obtained from 110 individuals. The proposed framework involved data acquisition, image preprocessing, offline data augmentation and data balancing, followed by fine-tuning and training the pre-trained EfficientNet-B0 model with a new set of dense layers and a softmax activation as an output layer and Comparative analysis against other CNN architectures, including ResNet-50, Xception, InceptionV3, VGG-16, and MobileNetV2. Experimental findings show that the proposed fine-tuned EfficientNet-B0 achieved a comparatively higher accuracy of 97.67%, outperforming the comparative studies. These results demonstrate the strong potential of the proposed network as an automated diagnostic tool in medical imaging. However, the limited dataset size represents a constraint, which may affect generalization to unseen data. Future work should focus on enlarging the training data via instances from real fields, applying more advanced image augmentation strategies, adjusting hyper parameters, and exploring novel architectures including U-Nets and Vision Transformers. Overall, this study validates the role of CNN-based approaches, particularly EfficientNet-B0, for lung cancer identification and establishes a basis for future research in the area of medical imaging.