No abstract

The divergence between labeled training data and unlabeled testing data is a significant challenge for recent deep learning models. Unsupervised domain adaptation (UDA) attempts to solve such a problem. Recent works show that self-training is a powerful approach to UDA. However, existing methods have difficulty in balancing scalability and performance. In this paper, we propose an instance adaptive self-training framework for UDA on the task of semantic segmentation. To effectively improve the quality of pseudo-labels, we develop a novel pseudo-label generation strategy with an instance adaptive selector. Besides, we propose the region-guided regularization to smooth the pseudo-label region and sharpen the non-pseudo-label region. Our method is so concise and efficient that it is easy to be generalized to other unsupervised domain adaptation methods. Experiments on 'GTA5 to Cityscapes' and 'SYN-THIA to Cityscapes' demonstrate the superior performance of our approach compared with the state-of-the-art methods.

Glomeruli instance segmentation from pathologic images is a fundamental step in the automatic analysis of renal biopsies. Glomerular histologic manifestations vary widely among diseases and cases, and several special staining methods are necessary for pathologic diagnosis. A robust model is needed to segment and classify glomeruli with different staining methods and apply in cases with various glomerular pathologic changes. Herein, pathologic images from renal biopsy slides stained with three basic special staining methods were used to build the data sets. The snapshot group included 1970 glomeruli from 516 patients, and the whole-slide image group included 8665 glomeruli from 148 patients. Cascade Mask region-based convolutional neural net architecture was trained to detect, classify, and segment glomeruli into three categories: i) GN, structural normal; ii) global sclerosis; and iii) glomerular with other lesions. In the snapshot group, total glomeruli, GN, global sclerosis, and glomerular with other lesions achieved an F1 score of 0.914, 0.896, 0.681, and 0.756, respectively, which were comparable with those in the whole-slide image group (0.940, 0.839, 0.806, and 0.753, respectively). Among the three categories, GN achieved the best instance segmentation effect in both groups, as determined by average precision, average recall, F1 score, and Mask mean Intersection over Union. The present model segments and classifies multistained glomeruli with efficiency and robustness. It can be applied as the first step for more detailed glomerular histologic analysis.

No abstract

No abstract