Search for: All records

We present a solution to image-based cell counting with dot annotations for both 2D and 3D cases. Current approaches have two major limitations: 1) inability to provide precise locations when cells overlap; and 2) reliance on costly labeled data. To address these two issues, we first adopt the inverse distance kernel, which yields separable density maps for better localization. Second, we take advantage of unlabeled data by self-supervised learning with focal consistency loss, which we propose for our pixel-wise task. These two contributions complement each other. Together, our framework compares favorably against stateof- the-art methods, including methods using full annotations on 2D and 3D benchmarks, while significantly reducing the amount of labeled data needed for training. In addition, we provide a tool to expedite the labeling process for dot annotations. Finally, we make the source code and labeling tool publicly available. 

Segmentation of echocardiograms plays an essential role in the quantitative analysis of the heart and helps diagnose cardiac diseases. In the recent decade, deep learning-based approaches have significantly improved the performance of echocardiogram segmentation. Most deep learning-based methods assume that the image to be processed is rectangular in shape. However, typically echocardiogram images are formed within a sector of a circle, with a significant region in the overall rectangular image where there is no data, a result of the ultrasound imaging methodology. This large non-imaging region can influence the training of deep neural networks. In this paper, we propose to use polar transformation to help train deep learning algorithms. Using the r-θ transformation, a significant portion of the non-imaging background is removed, allowing the neural network to focus on the heart image. The segmentation model is trained on both x-y and r-θ images. During inference, the predictions from the x-y and r-θ images are combined using max-voting. We verify the efficacy of our method on the CAMUS dataset with a variety of segmentation networks, encoder networks, and loss functions. The experimental results demonstrate the effectiveness and versatility of our proposed method for improving the segmentation results. 

Image-based cell counting is a fundamental yet challenging task with wide applications in biological research. In this paper, we propose a novel unified deep network framework designed to solve this problem for various cell types in both 2D and 3D images. Specifically, we first propose SAU-Net for cell counting by extending the segmentation network U-Net with a Self-Attention module. Second, we design an extension of Batch Normalization (BN) to facilitate the training process for small datasets. In addition, a new 3D benchmark dataset based on the existing mouse blastocyst (MBC) dataset is developed and released to the community. Our SAU-Net achieves state-of-the-art results on four benchmark 2D datasets - synthetic fluorescence microscopy (VGG) dataset, Modified Bone Marrow (MBM) dataset, human subcutaneous adipose tissue (ADI) dataset, and Dublin Cell Counting (DCC) dataset, and the new 3D dataset, MBC. The BN extension is validated using extensive experiments on the 2D datasets, since GPU memory constraints preclude use of 3D datasets. The source code is available at https://github.com/mzlr/sau-net. 

There is considerable interest in AI systems that can assist a cardiologist to diagnose echocardiograms, and can also be used to train residents in classifying echocardiograms. Prior work has focused on the analysis of a single frame. Classifying echocardiograms at the video-level is challenging due to intra-frame and inter-frame noise. We propose a two-stream deep network which learns from the spatial context and optical flow for the classification of echocardiography videos. Each stream contains two parts: a Convolutional Neural Network (CNN) for spatial features and a bi-directional Long Short-Term Memory (LSTM) network with Attention for temporal. The features from these two streams are fused for classification. We verify our experimental results on a dataset of 170 (80 normal and 90 abnormal) videos that have been manually labeled by trained cardiologists. Our method provides an overall accuracy of 91:18%, with a sensitivity of 94:11% and a specificity of 88:24%.