Two-stage training for abdominal pan-cancer segmentation in weak label

Hanwen Zhang, Yongzhi Huang, Bingding Huang

Keywords: Weak label, Pseudo label, Two-stage training

Abstract: Constructing comprehensive labeled datasets for medical image segmentation tasks is time-consuming, requiring intensive masks annotated carefully by experienced radiologists. Existing benchmark datasets provide the necessary masks to train the supervised-based segmentation models, including single-organ datasets and multiple-organ datasets. However, it is still challenging when deploying large-scale models with a union of multiple datasets due to annotation conflicts. For example, some organ or tumor annotations are missing in most cases (weak label) in the FLARE23 challenge dataset.To overcome the limitation of segmentation models in this situation, we propose a two-stage training method to train an efficient segmentation model with weak label. In the first stage, only strong labels (complete organ labels) are used to train models by the nnU-Net, while the weak labels (incomplete organ labels) are filled by generating pseudo labels using nnU-Net. Then the lightweight coarse-to-fine network is trained using the supplemented data in the second stage.Experiments on the FLARE23 challenge (MICCAI FLARE23) demonstrate that coarse-to-fine networks can reduce computational complexity and resource consumption during the inference stage while maintaining high performance, in the case of pseudo labeled supplementary data. With a speed of 12.6 seconds per case, our proposed method achieves an average DSC of 0.8920 and an average NSD of 0.9482 on the FLARE23 validation set.

Submission Number: 2