Multi-Organ and Pan-cancer Segmentation Framework from Partially Labeled Abdominal CT Datasets: Fine and Swift nnU-Nets with Label Fusion

Youngbin Kong, Kwangtai Kim, Seoi Jeong, Kyu Eun Lee, Hyoun-Joong Kong

Keywords: Label Fusion, Abdominal CT, Segmentation, Partially Labeled Dataset, Deep Learning

Abstract: Segmentation of organs and tumors from abdominal computed tomography (CT) scans is crucial for cancer diagnosis and surgical planning. Since traditional segmentation methods are subjective and labor-intensive, deep learning-based approaches have been introduced recently which incur high computational costs. This study proposes an accurate and efficient segmentation method for abdominal organs and tumors in CT images utilizing a partially-labeled abdominal CT dataset. Fine nnU-Net was used for the pseudo-labeling of unlabeled images. And the Label Fusion algorithm combined the benefits of the provided datasets to build an optimal training dataset, using Swift nnU-Net for efficient inference. In online validation using Swift nnU-Net, the dice similarity coefficient (DSC) values for organs and tumors segmentation were 89.56\% and 35.70\%, respectively, and the normalized surface distance (NSD) values were 94.67\% and 25.52\%. In our own efficiency experiments, the inference time was an average of 10.7 seconds and the area under the GPU memory time curve was an average of 20316.72. Our method enables accurate and efficient segmentation of abdominal organs and tumors using partially labeled data, unlabeled data, and pseudo-labels. This method could be applied to multi-organ and pan-cancer segmentation in abdominal CT images under low-resource environments.

Submission Number: 37