Teeth image segmentation has significant implications for various aspects of dental medicine, forensic odontology identification, scientific research, and education. However, the annotations made by radiologists may be subjective, and manual annotation requires a considerable amount of time and labor costs. In this paper, we propose deep-learning model VCMix-Net+ to achieve high-quality segmentation of 2D teeth images. Our VCMix-Net+ performs parallel operations of convolution and a multi-layer perceptron with class attention at multiple scales, simultaneously capturing both local and local-global information. Additionally, we have successfully reduced the annotation cost by employing a self-training semi-supervised approach. Our contribution lies in introducing a dedicated pseudo-label filtering network during the pseudo-label generation phase of semi-supervised training. This filtering network helps in selecting high-quality pseudo-labels, thereby reducing the risk of error accumulation in the self-training process. Meanwhile, we propose the concept of half-image training, which effectively improves the segmentation evaluation metrics. Finally, the average Score of our method’s predicted results on the online test set is 92.89.