Ann Arbor Pioneer High School,
Keywords: kidney cancer staging, AI, deep learning, CT
Summary:The American Cancer Society’s most recent estimates for kidney cancer in the United States for 2020 are that about 73,750 new cases of kidney cancer (45,520 in men and 28,230 in women) will occur, and about 14,830 people (9,860 men and 4,970 women) will die from this disease. In kidney cancer, stage 1 is an important threshold for the decision of organ preservation surgery verses chemotherapy and organ removal for higher stages. Incorrect staging results in under- or over-treatment. The purpose of this project is to design an Artificial Intelligence (AI) based system that will allow for a more accurate kidney cancer stage assessment by physicians. Deep Learning Neural Network (DLNN) was used to predict kidney cancer stage 1 verses higher stages. The publicly available cancer research database from the National Cancer Institute TCIA provided anonymized computer tomography (CT) scans and clinical data from 227 patients with kidney cancer of different stages used for training and testing of the system. The kidneys containing cancer were cropped from the 3D CT scans for all 227 patients using image editing software. Approximately 4000 cropped CT kidney images were obtained. The use of the cropped images had focused the Deep Learning Neural Network training only on information related to kidney cancer. Inception V3 deep learning network structure was used in this project within the TensorFlow platform. Transfer learning technique was used to train the Inception V3 for the task of staging kidney cancer. Transfer learning is useful when the dataset for the target task is not very large as it retrains only part of the network structure with the target task at hand. The Inception V3 was first trained using the ImageNet dataset, which consists of more than 1,000,000 natural scene images. Then only a part of the Inception V3 was retrained with the kidney cancer data. The input to the Inception V3 was the CT cropped image of kidney with cancer, where the output was the likelihood of the kidney cancer being Stage 1. The CT dataset was split into 40% training, 10% validation, and 50% testing sets. The DLNN was trained using the training and the validation sets and tested using the test set to ensure accurate and reliable evaluation results. The Deep learning neural network was trained up to 5000 iterations on the training set until the training error converged. The best DLNN was selected using the validation set. Then this model was used to run on the test set. The classification accuracy was estimated using the area under the ROC curve (AUC). The AUC of 0.96 for training, 0.88 for validation and 0.87 for the test sets was obtained. This AI system shows promise for potentially assisting physicians in kidney cancer staging.