Team: Gordon Center for Medical Imaging, Massachusetts General Hospital, Harvard Medical School, USA
Authors: Aoxiao Zhong, Quanzheng Li
A fully convolutional ResNet-101 network with atrous convolution and atrous spatial pyramid pooling was trained. Atrous convolution and atrous spatial pyramid pooling enlarge the field-of-view for prediction and allow capturing objects as well as image context at multiple scales. Patches of size 512*512 at level 1 were used as input to the network. For each mini-batch 10 normal and 10 tumor patches were randomly taken to train the network. The model was trained for 40000 iterations on a single P100 GPU. The classification task is based on features extracted from tumor probability map with a random forest classifier as described in the submission by HMS-MIT Method 1. Further postprocessing step was used to remove candidates that are too small for the lesion-detection task.
The following figure shows the receiver operating characteristic (ROC) curve of the method.
The following figure shows the free-response receiver operating characteristic (FROC) curve of the method.
The table below presents the average sensitivity of the developed system at 6 predefined false positive rates: 1/4, 1/2, 1, 2, 4, and 8 FPs per whole slide image.