Publications

Salient segmentation is a critical step in biomedical image analysis, aiming to cut out regions that are most interesting to humans. Recently, supervised methods have achieved promising results in biomedical areas, but they depend on annotated training data sets, which requires labor and proficiency in related background knowledge. In contrast, unsupervised learning makes data-driven decisions by obtaining insights directly from the data themselves. In this paper, we propose a completely unsupervised self- aware network based on pre-training and attentional backpropagation for biomedical salient segmentation, named as PUB-SalNet. Firstly, we aggregate a new biomedical data set from several simulated Cellular Electron Cryo- Tomography (CECT) data sets featuring rich salient objects, different SNR settings and various resolutions, which is called SalSeg-CECT. Based on the SalSeg-CECT data set, we then pre-train a model specially designed for biomedical tasks as a backbone module to initialize network parameters. Next, we present a U-SalNet network to learn to selectively attend to salient objects. It includes two types of attention modules to facilitate learning saliency through global contrast and local similarity. Lastly, we jointly refine the salient regions together with feature representations from U-SalNet, with the parameters updated by self- aware attentional backpropagation. We apply PUB-SalNet for analysis of 2D CECT images and achieve state-of-the-art performance on simulated biomedical data sets. Furthermore, our proposed PUB-SalNet can be easily extended to 3D images. The experimental results on the 2d and 3d data sets also demonstrate the generalization ability and robustness of our method.

Pre-trained networks have recently achieved great success in computer vision. At present, most deep learning-based saliency detection methods use pre-trained networks to extract features, regardless of supervised or unsupervised. However, we found that when unsupervised saliency detection is performed on grayscale biomedical images, pre-trained networks such as VGG cannot effectively extract significant features. We suggest that VGG is not able to learn salient information from grayscale biomedical images and its performance greatly depends on RGB cues and quality of the training set. To verify our hypothesis, we construct an adversarial data set featuring a low signal-to-noise ratio (SNR), low resolution and rich salient objects and conduct a series of probing experiments. What’s more, in order to further explore what VGG has learned, we visualize intermediate feature maps. To the best of our knowledge, we are the first to investigate the reliability of deep learning methods for unsupervised saliency detection on grayscale biomedical images. It’s worth noticing that our adversarial data set also provides a more robust evaluation of saliency detection and may serve as a standard benchmark in future work on this task.