Program Information
Automatic Lung Cancer Detection From CT Using a GPU-Accelerated Deep Convolutional Neural Networks
H Lin1*, T Liu1 , C Shi2 , X Tang3 , X Pei4 , X Xu1,4 (1) Rensselaer Polytechnic Institute, Troy, NY (2) Memorial Sloan Kettering Cancer Center, Basking Ridge, NJ (3) Memorial Sloan Kettering Cancer Center, West Harrison, NY (4) University of Science and Technology of China, Hefei, Anhui
Presentations
TU-H-CAMPUS-JT-1 (Tuesday, August 1, 2017) 4:30 PM - 5:30 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: Clinical challenges exist for accurate diagnosis of lung cancer. Deep Convolutional Neural Networks (CNN) is a powerful tool for computer vision tasks, and is promising to lead to improved accuracy of medical image analysis. We herein propose a GPU-accelerated 3D CNN classifier to detect cancerous lesions in the lungs accurately and fast, providing decision support to the clinical diagnostic workflow.
Methods: The dataset includes 1397 thoracic CT image sets with annotations provided by radiologists. Each CT set was pre-processed into 50x50 resolution and rendered into 20 slices in compliance with the need for data uniformity and limited memory of the GPUs. Our CNN consists of two convolutional layers, max-pooling layers, fully-connected layers, a dropout layer, and a final softmax layer for lesion classification. The patch size of two convolutional layers is 3x3x3 and the stride size is 1. Max-pooling layer comes with a window size of 2x2x2 with stride size of 2. Stochastic gradient descent method was used to minimize the classification loss function. The CNN was built with the TensorFlow framework and trained on a NVidia Titan X Pascal GPU. The CNN was trained for 100 epochs to obtain the classification accuracy and cost.
Results: A 3D CNN was developed for the lung cancer detection and trained on GPUs. An overall classification accuracy of 80.2% over 10-fold cross validation has been achieved, improving the accuracy of published methods by 5%-12%. For performance, the training time using a Titan X GPU is 52 minutes, which is over 30 times faster than a 12-threads CPU.
Conclusion: The proposed 3D CNN classifier for lung cancer detection has demonstrated the feasibility to enhance the decision support to the clinical diagnosis. We are expanding the dataset and fine-tuning the CNN parameters and more statistically significant results are to be presented.
Contact Email: