Pneumonia infection segmentation using semi-supervised learning and uncertainty estimation

Accurate segmentation of pneumonia lesions is critical for early diagnosis and clinical treatment, while traditional supervised learning methods depend on a substantial amount of labeled data, which are costly to acquire. To address the above issues, this work guides segmentation network training based on semi-supervised learning using an uncertainty estimation that makes full use of unlabeled data to simultaneously extract CNN features from input 3D images at multiple dimensional scales, as well as the use of consistency loss to constrain the consistency of multiscale predictions within and across models. Furthermore, in order to enhance the dependability of model predictions for unlabeled data, the segmentation training was guided by constructing an anatomically aware and multiple dimensional-attention (AA-MDA) network and calculating an uncertainty map. First, unreliable segmentation predictions are fed into an AA-MDA network that has been pre-trained and has an anatomically-aware prior, which subsequently adapts the unreliable predictions to reasonable segmentations through pre-learned representations, and semi-supervised training is guided by comparing the difference between the actual segmentation result and the reasonable segmentation as an uncertainty estimate. Ultimately, the segmentation network achieved high accuracy on the MosMedData dataset.