Deep learning methods have been extensively employed in medical image analysis tasks, and displayed a considerable improvement in performance and precision. Nonetheless, these methods are frequently impeded by the requirement for a substantial amount of annotated data, which is often expensive and time-consuming to acquire, particularly for medical images. Recently, the incorporation of unlabeled data into limited labeled data through semi-supervised learning has alleviated this problem to some extent, exhibiting promising outcomes on various image analysis tasks. Fixmatch, developed by the Google research team, is a commonly implemented hybrid semi-supervised learning approach that delivers exceptional outcomes in multiple image classification tasks. This study presents PixelFixmatch as an extension of Fixmatch for medical image segmentation. The proposed approach incorporates pixel attention during the generation of pseudo labels. The weak and strong data augmentation processes are inherited, and the prediction of weakly augmented unlabeled data is used to supervise strongly augmented data. High-quality pseudo labels at the pixel level are obtained by analyzing the predicted heatmap of the weakly augmented data. The loss is then weighted by pixel-wise confidence, enabling the extraction of useful information and the exclusion of ambiguous data. Our approach exhibits significant advantages over the powerful nnU-Net benchmark on both the public dataset Medical Segmentation Decathlon and two private medical image datasets. We conducted ablation studies to distinguish and confirm the effectiveness of key designs.