Pseudo Label-Based Semi-Supervised Learning for Abdominal Organ and Cancer Segmentation in CT Image With Partial Labeled Data

Panlong Xu, Zhijian Li, Weiping Liu

Keywords: emi-supervised learning · Image segmentation · Pseudo label.

TL;DR: We propose a semi-supervised learning approach based on nnUNet to solve the abdominal multi-organ and pan-tumor segmentation problem in CT images.

Abstract:

Abdominal multi-organ and pan-tumor segmentation in CT
image plays a critically important role in preoperative planning, intraoperative
navigation, and postoperative assessment for surgical procedures.
In this study, we propose a semi-supervised learning approach using
nnU-Net on the FLARE2023 competition dataset. Our methodology
involves training an initial model on fully annotated data, followed
by inference on partially annotated data to generate pseudo-labels, and
subsequently training a final model using these pseudo-labeled data. To
optimize computational efficiency, we adopt a parameter-efficient model
with a reduced number of parameters. By leveraging the availability of
both labeled and unlabeled data, our approach aims to enhance the performance
of the nnU-Net model while maintaining a reasonable computational
cost. Ultimately, our trained small nnU-Net achieved significant
results on a validation set of 100 samples, with a dice coefficient of 0.8854
for multi-organ segmentation and 0.4186 for tumor segmentation. Moreover,
the average inference time of the model was only 18 seconds.

Submission Number: 34