Program Information
Deep Convolutional Neural Networks for Prostate Cancer Detection On Multiparametric MRI : A Transfer Learning Approach
Q Chen1*, S Hu2 , X Xu2 , X Li3 , Q Zou3 , Y Li2 , (1) University of Virginia, Charlottesville, VA, (2) SkyData InfoTech, Nanjing, Jiangsu, (3) Affiliated Cancer Hospital of Nanjing Medical University, Nanjing, Jiangsu
Presentations
TU-FG-605-2 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: To automatically assess malignancy of prostate lesions based on multi-parametric magnetic resonance images (mpMRI).
Methods: A mpMRI prostate dataset consisting of 344 patient cases and 538 point of interests (POIs) was used. Each POI was at the center of a suspected lesion. The lesion’s malignancy status was established by biopsy. The prostate zone for each lesion was also given. 203 cases was allocated for training, the remaining cases was used as test dataset to evaluate the performance. Two ImageNet pre-trained deep convolutional neural network (DCNN), Inception-V3 and Vgg-16, was re-trained on this training dataset with most of the layers frozen except for the final full-connection classification layer. Two approaches were used for validation: pre-allocate 1/3 of the training data and five-fold cross-validation. The area under the receiver operating characteristic curve (AUC) was used as performance metric. Re-scaling was applied to scores from different prostate zone to maximize the AUC for all lesions.
Results: Both models obtained an AUC of 0.86 on the training dataset. On test dataset, the Vgg-16 model obtained an AUC of 0.83 while the Inception-V3 obtained an AUC of 0.81. These performances are comparable to those obtained by expert radiologists. By proper rescaling, the AUC improved from 0.82 to 0.86 on the training dataset.
Conclusion: Good results were obtained with transfer learning to classify prostate lesions on mpMRI images. It shows that the computer-aided detection of prostate cancer using DCNN is a very promising direction. Training sophisticated modern DCNNs from scratch requires very large dataset and a long time, which are often difficult to obtain in most clinical applications. Transfer learning allows us to train those models with a small training dataset in a much shorter time while still achieving excellence performance.
Contact Email: