Deep Learning-Based Lung Cancer Detection and Classification Using CT Scan Image

Abstract

Early detection is significant for improving lung cancer patient outcomes. However, risk concerns and resource scarcity bring challenges, especially in developing nations. This study offers a method for precise and effective lung cancer detection utilizing deep learning techniques. The black-box nature of deep neural networks challenges its use in mission-critical applications, raising ethical and moral concerns inducing a lack of trust. Explainable Artificial Intelligence and Result-based visualization techniques are methods used to add clinical trust to the decision logic of a deep learning models. In this study, four Pre-trained deep learning models are utilized for feature extraction from computed tomography images, optimizing computing complexity. Lung CT data was gathered from Tikur Anbessa Specialized Hospital in Addis Ababa, Ethiopia, to validate the proposed approach. After data cleaning is applied, 3500 CT images are prepared where 2000 images are cancerous and 1500 images are normal. Data augmentation is employed to improve the training performance by increasing training data size. For both classes, 90% of the dataset is used for training, while 10% is used for validation. In this study, we used GradCAM and LIME to visualize the detection’s region of interest and detection logic. GradCAM is a visualization tool that demonstrates special features that lead a convolutional neural network to generate its decisions on a target image, while LIME is a technique that explains the predictions of any classifier by approximating it locally with an interpretable model, thereby providing insights into the specific features influencing the model's decisions. Utilizing GradCAM and LIME, our study achieved a promising result in visualizing and comprehending the features that affect the model's detection, overcoming the black-box nature of deep learning. The outcomes of this study offer an opportunity to substantially improve the detection of lung cancer. As a result, the outperforming Densenet169 obtained 94.85% accuracy, 93.42% precision, 94.67% recall, and a 94.04% F1 score. As future work, we plan to improve data size for robust model training and better performance.