Self-training with Selective Re-training Improves Abdominal Organ Segmentation in CT Image

Fan Zhang, Meihuan Wang, Hua Yang

Keywords: Self-training · Efficient segmentation learning · Abdominal organ segmentation.

Abstract: Abstract. Inspired by self-training learning via pseudo labeling, we construct self-training framework with selective re-training pseudo labels to improve semi-supervised abdominal organ segmentation. In this work, we carefully design the strong data augmentations (SDA) and test-time augmentations (TTA) to alleviate overfitting noisy labels as well as decouple similar predictions between the teacher and student models. For efficient segmentation learning (ESL), knowledge distillation is adopted to transfer larger teacher model to smaller student model for compressing model. We also develop the post half precision (FP16) quantization to accelerate the model inference. In addition, we propose the singlelabel based connected component labelling (CCL) for post processing. Compared to one-hot CCL of O(n) time complexity, which on the singlelabel based method is reduced to O(1). Quantitative evaluation on the FLARE2022 validation cases, this method achieves the average dice similarity coefficient (DSC) of 0.8813 on semi-supervised model, it achieves remarkable improvement compared to 0.7711 on full-supervised model. Code is available at https://github.com/Shanghai-Aitrox-Technology/EfficientSegLearning