Using Transfer Learning to classify different subtypes of non-small cell lung cancer
Last year, over 200,000 people in America were diagnosed with lung cancer. Current methods for diagnosing lung cancer are inaccurate and often are too late to help. There exists both a high positive and high negative rate, making it difficult to determine whether someone actually has lung cancer. Improving lung cancer diagnosis could potentially save thousands of lives. In this paper, I propose a transfer learning approach to differentiate between different subtypes of non-small cell lung cancer(NSCLC), using CT-scans. I transfer learned my data onto the InceptionV3 model, which was pretrained on the ImageNet dataset. I feed in each individual slice of a CT-scan into a deep learning model, which then classifies the slice as a subtype of NSCLC. The model has a 98% train accuracy and a 85% validation accuracy performance on 317 patients who had various subtypes of NSCLC. On the test dataset, it obtains an accuracy of 78% and a precision of 79.4%. The model achieved a AUC-ROC of 89.6% using macro averaging, and 91.2% on micro averaging. This demonstrates the potential of transfer learning for 3D medical images, as well as a possibility for deep learning to improve the accuracy and efficiency of lung cancer diagnosis.
This code is for classifying adenocarcinoma, squamous cell carcinoma, and large cell carcinoma from CT-scans of patients with Lung Cancer. The data can all be found from the NSCLC Radiomics dataset, which can be found here: https://wiki.cancerimagingarchive.net/display/Public/NSCLC-Radiomics