Selected Partially Labeled Learning for Abdominal Organ and Pan-cancer Segmentation

Yuntao Zhu

Keywords: Partially labeled learning, Accelerate inference, Lightweight network

Abstract:

Obtaining labeled data from medical images is very expensive and labor intensive. At the same time, the large number of existing publicly available medical image datasets are usually labeled with only some of the organs as target regions, while other organs in the image are ignored. It is a challenge to train a neural network to segment all labeled categories using only partially labeled data.
We design a compound loss, the selected partially cross entropy and dice loss, that allows the neural network to learn specific categories from partially labeled data.
In addition, we improve the inference and training process of nnU-Net to reduce computational resources and accelerate inference.
Experiments demonstrate that our method achieves the average Dice Similarity Coefficient of 0.8514 and 0.1514 on 13 abdominal organ and tumor segmentation tasks, and enables the network to efficiently segment specific categories from partially labeled data.
Moreover, it significantly improves the inference speed, with an average running time of 21.8 seconds, and uses only an average of 2531 MB of maximum GPU memory.

Submission Number: 12