WITS: Weakly-supervised individual tooth segmentation model trained on box-level labels

•We propose an ellipses detection method which can better fit the shape of tooth.•We propose a level-set-based curvature model with restriction term which can force the active contour to fit the non-convex tooth and preserve the signed distance property of level function.•The proposed model is a weakly-supervised model which is trained on box-level labels but can obtain pixel-level results.