A Hybrid Deep Learning Framework for Early Detection of Lung Cancer Using CT Scan Imaging

Abstract

Lung cancer accounts for the most cancer deaths in the world. While CT scans can detect lung cancer at an early stage, the manual reading of these scans is time-consuming and varies from doctor to doctor. To overcome these challenges, a framework is proposed based on deep learning techniques to classify CT chest scans as normal or cancerous. Two deep CNN architectures, namely VGG16 and ResNet50, were used, both pretrained on the ImageNet dataset. These models were trained to train a compact ANN. Six different models were trained using these architectures and three randomly chosen seeds. The best performing model on the validation set was selected and used for classification of the test data after applying test time augmentations. The model was trained on the Kaggle Chest CT-Scan Images dataset, which contains 1,000 CT scan images, divided into 80%, 10%, and 10% splits for training, validation, and test data, respectively. The proposed framework achieved accuracy, sensitivity, specificity, and area under the ROC curve values of 99.0%, 100%, 95.5%, and 0.999, respectively. Further metrics evaluated for this model included the Matthews Correlation Coefficient, balanced accuracy, area under the precision-recall curve, calibration of the model, 95% confidence intervals, and a comparison to four other deep learning frameworks for classification of the same type of CT scans. The deep learning framework proposed for this study was found to be statistically equivalent to the other deep learning frameworks (p ≥ 0.5 for all models evaluated) yet utilized the fewest number of parameters (14.85 M). Furthermore, an analysis of the computational complexity and robustness of the model against perturbations to the CT scan acquisition was presented. Finally, another evaluation of the framework was performed on a different, independent dataset (IQ-OTH/NCCD). For this separate dataset, the model achieved 99.4% accuracy on the test data after being retrained on the new data. A model trained on the Kaggle Chest CT-Scan Images dataset, however, only achieved 64.7% accuracy for its predictions on scans from this other independent dataset. These results indicate that the deep learning framework was able to generalize to new data, but the trained deep learning model did not generalize to scans from new institutions. Thus, this deep learning framework was able to achieve high accuracy with few parameters yet remain robust in various evaluation conditions.