Using Deep Convolutional Neural Networks to Optimize Pulmonary Nodule Classification and Localization in Radiograph Imaging for Early Lung Cancer Detection
By James Wang
Preliminary diagnosis of lung cancer has led to countless cases of overtreatment due to false positive classifications made by physicians and radiologists. Most commonly, the misclassification of a benign pulmonary nodule (PN) as malignant from chest X-ray images initiates this process for patients. In the advent of promising machine learning and computer vision models, we investigate the optimization of benign and malignant PN classification using deep convolutional neural networks through transfer learning by fine tuning its convolutional layers. Specifically, we look at how fine-tuning the VGG19 convolutional neural network model differently affects its classification accuracy. With our optimal model, we test its efficacy in localizing and classifying PNs on chest radiographs using a selection search-based scanning method. We found that fine-tuning the last convolutional block yields the highest predictive performance. Using a reserved image test set, our model is able to yield a classification accuracy of 77% compared to published models yielding 68%. This methodology can be easily generalized and applied to other medical imaging tasks.