Ipnet: informative patches learning for semi-supervised magnetic resonance image segmentation.

Semi-supervised learning has become a favorable method for medical image segmentation due to the high cost of obtaining labeled data in the field of medical image analysis. However, existing magnetic resonance images have low contrast, the scale and shape of organs vary greatly under different slice perspectives. Although existing methods have made some progress, they still cannot handle these challenging samples well. To this end, we propose a semi-supervised magnetic resonance images segmentation method based on informative patches learning (IPNet), which focuses on the learning of challenging regions. Specifically, we design a novel informative patch scoring strategy based on prediction uncertainty and category diversity, which can accurately identify challenging regions in samples. And to ensure that the informative patch is fully learned, the patch with the lowest score in one sample is replaced with the patch with the highest score in another sample to obtain a new pair of training samples. Furthermore, we introduce global and local consistency losses to supervise the new samples, guide the model to focus on the global and local features of the informative patches. To evaluate the effectiveness of the method, we conducted experiments on three magnetic resonance image datasets (ACDC, PROMISE 12 and LA datasets). Extensive experimental results demonstrate the effectiveness and superior performance of the proposed method.