DenseNet201-Based Deep Transfer Learning Framework for Accurate Multi-Class Lung Cancer Classification Using Computed Tomography Images

Abstract

Timely and precise diagnosis of lung cancer is vital to better survival and effective clinical decision making. With the advent of deep learning, automated medical image analysis has shown great promise; however, simultaneously attaining high classification accuracy, robustness, and interpretability of the model is difficult. In this study, a Deep Transfer Learning (DTL) approach with DenseNet201 network is proposed for classification of lung cancer if it is multi-class in nature, based on the computed tomography (CT) images.In this study, a DenseNet201 based deep transfer learning (DTL) framework is proposed for multi-class lung cancer classification using computed tomography (CT) images from publicly available IQ-OTHNCCD dataset. The proposed architecture consists of a DenseNet201 backbone pre-trained with ImageNet and a custom classification head (GAP, BN, FC, Dropout, Softmax). A comprehensive data augmentation, label smoothing, class-weight balancing, early stopping, adaptive learning rate scheduling, and two-stage fine-tuning were used in the training process for improving generalization and preventing overfitting. The CT images of the dataset were divided into benign, malignant and normal classes, with 767, 164 and 166 images in the training, validation, and testing sets respectively. Experimental results showed that DenseNet201 framework attained the highest accuracy of 96.39%, precision of 96.49%, recall of 96.39%, and F1-score of 96.30%, in the test phase. The proposed model has been compared with ResNet50 (87.95%) and EfficientNetB0 (72.89%) under the same experimental conditions, and it has been confirmed that the proposed model has a stronger feature extraction ability for lung CT image classification. In addition, the classification accuracy was validated by a confusion matrix analysis, receiver operating characteristic (ROC) curves, area under the curve (AUC), and class-wise sensitivity and specificity analysis were performed, and the model interpretability was increased by using the Gradient-weighted Class Activation Mapping (Grad-CAM). The experimental results show that the proposed DenseNet201 framework is able to classify lung cancers accurately, robustly and explainably into multi-class and has great potential for use in computer-aided diagnosis (CAD) system for early detection of lung cancer by radiologists.