Xiaoyang Han1, Yuting Zhai1, Botao Zhao1, and Xiao-Yong Zhang1
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China
1Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University, Shanghai, China
We develop a multi-task learning model for extreme small lesion detection tasks to detect magnetic particle deposition in MR images. The proposed end-to-end network can simultaneously detect the number and the location of small lesions with acceptable precision and sensitivity.
Figure 1. Multi-task learning structure. The input images are on the left; the output count of lesions is on the bottom right; the output map is on the top right: distance map for point annotation dataset and segmentation map for full segmentation dataset. After each convolutional layer, a rectified linear unit activation is applied. In the regression network, after the last convolutional layer, a global max-pooling layer is added to connect the following fully connected layers.
Figure 4. Representative detection results. In the prediction image (right column): green circle for true positive, red circle for false positive and orange circle for false negative.