A Lightweight nnU-Net Combined with Target Adaptive Loss for Organs and Tumors Segmentation

Accurate and automated abdominal organs and tumors segmentation is of great importance in clinical practice. Due to the high time- and labor-consumption of manual annotating datasets, especially in the highly specialized medical domain, partially annotated datasets and unlabeled datasets are more common in practical applications, compared to fully labeled datasets. CNNs based methods have contributed to the development of medical image segmentation. However, previous CNN models were mostly trained on fully labeled datasets. So it is more vital to develop a method based on partially labeled datasets. In FLARE23, we design a model combining a lightweight nnU-Net and target adaptive loss (TAL) to obtain the segmentation results efficiently and make full use of partially labeled dataset. Our method achieved an average DSC score of 86.40% and 19.41% for the organs and lesions on the validation set and the average running time and area under GPU memory-time cure are 25.34 s and 23018 MB, respectively.