Partial annotation-based organs and tumor segmentation with progressive weakly supervised learning

Dengqiang Jia, Zilong Wang

Keywords: Partial label, Self-training, Organ segmentation

Abstract:

In medical image analysis, obtaining labeled data is expensive and time-consuming.
Numerous unlabeled data can be used for efficient abdominal organ segmentation.
Besides, partially annotated data is easier to collect and can be used to develop label-efficient algorithms, reducing the annotation cost for considerable performance.
We proposed a progressive weakly supervised learning for abdomen organs and tumor segmentation, i.e., PWS-Seg.
PWS-seg can learn from organs to tumors via a progressive framework based on partially annotated images.
Moreover, we applied a class-wise label fusion strategy to get a new set of reliable pseudo labels.
On the FLARE2023 online validation cases, with the help of unlabeled data, our method obtained the average dice similarity coefficient (DSC) of 82.68% and average normalized surface distance (NSD) of 86.00%, which is better than the method only using partial annotated.
The average running time is 100.17s per case in the inference phase, and the maximum used GPU
memory is 4128 MB.

Submission Number: 3