Unsupervised body part regression using convolutional neural network with self-organization
share
Automatic body part recognition for CT slices can benefit various medical image applications. Existing methods require large-scale datasets of labeled images. However, the intrinsic structure information in CT volumes was still not fully leveraged. In this paper, we propose unsupervised body part regressor (UBR) to utilize this information. A novel learning framework based on a convolutional neural network and two inter-sample loss functions was designed. UBR builds a coordinate system for the body and outputs a continuous score for each slice, which represents the normalized position of the body part in the slice. The training process of UBR resembles a self-organization process: slice scores are learned from inter-slice relationships. The training samples are unlabeled volumes that are abundant in every hospital's database, thus zero annotation effort is needed. UBR is simple, fast, and accurate. Qualitative and quantitative experiments proved its effectiveness. Besides, we show potential applications of UBR in network initialization and anomaly detection.
READ FULL TEXT