Yibiao Rong

Convolutional neural networks (CNNs) are effective tools for regression tasks. However, their black-box nature limits their applicability in high-impact and high-risk tasks. In this paper, a novel method is proposed to identify particular patterns in an image that can make the output of a CNN model equal to a specified value, thereby helping users understand the behaviours of CNNs. Specifically, in the proposed method, a set of binary filters is first randomly initialized. A genetic algorithm is then employed to evolve the binary filters such that the output of the CNN is equal to a specified value when taking a filtered image, which is obtained by convolving an original image and an evolved filter, as its input. Many experiments are conducted to evaluate the effectiveness of the proposed method. The results show that the proposed method is highly effective at identifying the patterns that can make a CNN output a specified value. The codes of this paper are available at https://github.com/ybrong-stu/Identifying. git

Deep convolutional neural networks (CNNs) have been widely used for fundus image classification and have achieved very impressive performance. However, the explainability of CNNs is poor because of their black-box nature, which limits their application in clinical practice. In this paper, we propose a novel method to search for discriminative regions to increase the confidence of CNNs in the classification of features in specific category, thereby helping users understand which regions in an image are important for a CNN to make a particular prediction. In the proposed method, a set of superpixels is selected in an evolutionary process, such that discriminative regions can be found automatically. Many experiments are conducted to verify the effectiveness of the proposed method. The average drop and average increase obtained with the proposed method are 0 and 77.8%, respectively, in fundus image classification, indicating that the proposed method is very effective in identifying discriminative regions. Additionally, several interesting findings are reported: 1) Some superpixels, which contain the evidence used by humans to make a certain decision in practice, can be identified as discriminative regions via the proposed method; 2) The superpixels identified as discriminative regions are distributed in different locations in an image rather than focusing on regions with a specific instance; and 3) The number of discriminative superpixels obtained via the proposed method is relatively small. In other words, a CNN model can employ a small portion of the pixels in an image to increase the confidence for a specific category.

Deformable image registration can obtain dynamic information about images, which is of great significance in medical image analysis. The unsupervised deep learning registration method can quickly achieve high registration accuracy without labels. However, these methods generally suffer from uncorrelated features, poor ability to register large deformations and details, and unnatural deformation fields. To address the issues above, we propose an unsupervised multi-scale correlation iterative registration network (SearchMorph). In the proposed network, we introduce a correlation layer to strengthen the relevance between features and construct a correlation pyramid to provide multi-scale relevance information for the network. We also design a deformation field iterator, which improves the ability of the model to register details and large deformations through the search module and GRU while ensuring that the deformation field is realistic. We use single-temporal brain MR images and multi-temporal echocardiographic sequences to evaluate the model's ability to register large deformations and details. The experimental results demonstrate that the method in this paper achieves the highest registration accuracy and the lowest folding point ratio using a short elapsed time to state-of-the-art.

Automatic and accurate estimation of choroidal thickness plays a very important role in a computer-aided system for eye diseases. One of the most common methods for automatic estimation of choroidal thickness is segmentation-based methods, in which the boundaries of the choroid are first detected from optical coherence tomography (OCT) images. The choroidal thickness is then computed based on the detected boundaries. A shortcoming in the segmentation-based methods is that the estimating precision greatly depends on the segmentation results. To avoid the dependence on the segmentation step, in this paper, we propose a direct method based on convolutional neural networks (CNNs) for estimating choroidal thickness without segmentation. Concretely, a B-scan image is first cropped into several patches. A trained CNN model is then used to estimate the choroidal thickness for each patch. The mean thickness of the choroid in the B-scan is obtained by taking the average of the choroidal thickness on each patch. Then, 150 OCT volumes are collected to evaluate the proposed method. The experiments show that the results obtained by the proposed method are very competitive with those obtained by segmentation-based methods, which indicates that direct estimation of choroidal thickness is very promising.

Recently, many methods based on hand-designed convolutional neural networks (CNNs) have achieved promising results in automatic retinal vessel segmentation. However, these CNNs remain constrained in capturing retinal vessels in complex fundus images. To improve their segmentation performance, these CNNs tend to have many parameters, which may lead to overfitting and high computational complexity. Moreover, the manual design of competitive CNNs is time-consuming and requires extensive empirical knowledge. Herein, a novel automated design method, called Genetic U-Net, is proposed to generate a U-shaped CNN that can achieve better retinal vessel segmentation but with fewer architecture-based parameters, thereby addressing the above issues. First, we devised a condensed but flexible search space based on a U-shaped encoder-decoder. Then, we used an improved genetic algorithm to identify better-performing architectures in the search space and investigated the possibility of finding a superior network architecture with fewer parameters. The experimental results show that the architecture obtained using the proposed method offered a superior performance with less than 1% of the number of the original U-Net parameters in particular and with significantly fewer parameters than other state-of-the-art models. Furthermore, through in-depth investigation of the experimental results, several effective operations and patterns of networks to generate superior retinal vessel segmentations were identified. The codes of this work are available at https://github.com/96jhwei/Genetic-U-Net .

In our previous work, by combining the Hilbert scan with the symbol grouping method, efficient run-length-based entropy coding was developed, and high-efficiency image compression algorithms based on the entropy coding were obtained. However, the 2-D Hilbert curves, which are a critical part of the above-mentioned entropy coding, are defined on squares with the side length being the powers of 2, i.e., 2ⁿ, while a subband is normally a rectangle of arbitrary sizes. It is not straightforward to modify the Hilbert curve from squares of side lengths of 2ⁿ to an arbitrary rectangle. In this short article, we provide the details of constructing the modified 2-D Hilbert curve of arbitrary rectangle sizes. Furthermore, we extend the method from a 2-D rectangle to a 3-D cuboid. The 3-D modified Hilbert curves are used in a novel 3-D transform video compression algorithm that employs the run-length-based entropy coding. Additionally, the modified 2-D and 3-D Hilbert curves introduced in this short article could be useful for some unknown applications in the future.

Background: Advanced prediction of the daily incidence of COVID-19 can aid policy making on the prevention of disease spread, which can profoundly affect people's livelihood. In previous studies, predictions were investigated for single or several countries and territories. Objective: We aimed to develop models that can be applied for real-time prediction of COVID-19 activity in all individual countries and territories worldwide. Methods: Data of the previous daily incidence and infoveillance data (search volume data via Google Trends) from 215 individual countries and territories were collected. A random forest regression algorithm was used to train models to predict the daily new confirmed cases 7 days ahead. Several methods were used to optimize the models, including clustering the countries and territories, selecting features according to the importance scores, performing multiple-step forecasting, and upgrading the models at regular intervals. The performance of the models was assessed using the mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient, and Spearman correlation coefficient. Results: Our models can accurately predict the daily new confirmed cases of COVID-19 in most countries and territories. Of the 215 countries and territories under study, 198 (92.1%) had MAEs 0.8. For the 215 countries and territories, the mean MAE was 5.42 (range 0.26-15.32), the mean RMSE was 9.27 (range 1.81-24.40), the mean Pearson correlation coefficient was 0.89 (range 0.08-0.99), and the mean Spearman correlation coefficient was 0.84 (range 0.2-1.00). Conclusions: By integrating previous incidence and Google Trends data, our machine learning algorithm was able to predict the incidence of COVID-19 in most individual countries and territories accurately 7 days ahead.

Retinal vessel segmentation is a crucial step in diagnosing and screening various diseases, including diabetes, ophthalmologic diseases, and cardiovascular diseases. In this paper, we propose an effective and efficient method for vessel segmentation in color fundus images using encoder-decoder based octave convolution networks. Compared with other convolution networks utilizing standard convolution for feature extraction, the proposed method utilizes octave convolutions and octave transposed convolutions for learning multiple-spatial-frequency features, thus can better capture retinal vasculatures with varying sizes and shapes. To provide the network the capability of learning how to decode multifrequency features, we extend octave convolution and propose a new operation named octave transposed convolution. A novel architecture of convolutional neural network, named as Octave UNet integrating both octave convolutions and octave transposed convolutions is proposed based on the encoder-decoder architecture of UNet, which can generate high resolution vessel segmentation in one single forward feeding without post-processing steps. Comprehensive experimental results demonstrate that the proposed Octave UNet outperforms the baseline UNet achieving better or comparable performance to the state-of-the-art methods with fast processing speed. Specifically, the proposed method achieves 0.9664 / 0.9713 / 0.9759 / 0.9698 accuracy, 0.8374 / 0.8664 / 0.8670 / 0.8076 sensitivity, 0.9790 / 0.9798 / 0.9840 / 0.9831 specificity, 0.8127 / 0.8191 / 0.8313 / 0.7963 F1 score, and 0.9835 / 0.9875 / 0.9905 / 0.9845 Area Under Receiver Operating Characteristic curve, on DRIVE, STARE, CHASE_DB1, and HRF datasets, respectively.