Stereo Matching Using Multi-Level Cost Volume and Multi-Scale Feature Constancy.

For CNNs based stereo matching methods, cost volumes play an important role in achieving good matching accuracy. In this paper, we present an end-to-end trainable convolution neural network to fully use cost volumes for stereo matching. Our network consists of three sub-modules, i.e., shared feature extraction, initial disparity estimation, and disparity refinement. Cost volumes are calculated at multiple levels using the shared features, and are used in both initial disparity estimation and disparity refinement sub-modules. To improve the efficiency of disparity refinement, multi-scale feature constancy is introduced to measure the correctness of the initial disparity in feature space. These sub-modules of our network are tightly-coupled, making it compact and easy to train. Moreover, we investigate the problem of developing a robust model to perform well across multiple datasets with different characteristics. We achieve this by introducing a two-stage finetuning scheme to gently transfer the model to target datasets. Specifically, in the first stage, the model is finetuned using both a large synthetic dataset and the target datasets with a relatively large learning rate, while in the second stage the model is trained using only the target datasets with a small learning rate. The proposed method is tested on several benchmarks including the Middlebury 2014, KITTI 2015, ETH3D 2017, and SceneFlow datasets. Experimental results show that our method achieves the state-of-the-art performance on all the datasets. The proposed method also won the 1st prize on the Stereo task of Robust Vision Challenge 2018.

Novel Convergence Results of Adaptive Stochastic Gradient Descents.

Adaptive stochastic gradient descent, which uses unbiased samples of the gradient with stepsizes chosen from the historical information, has been widely used to train neural networks for computer vision and pattern recognition tasks. This paper revisits the theoretical aspects of two classes of adaptive stochastic gradient descent methods, which contain several existing state-of-the-art schemes. We focus on the presentation of novel findings: In the general smooth case, the nonergodic convergence results are given, that is, the expectation of the gradients' norm rather than the minimum of past iterates is proved to converge; We also studied their performances under Polyak-Lojasiewicz property on the objective function. In this case, the nonergodic convergence rates are given for the expectation of the function values. Our findings show that more substantial restrictions on the steps are needed to guarantee the nonergodic function values' convergence (rates).

Nonergodic Complexity of Proximal Inertial Gradient Descents.

The proximal inertial gradient descent (PIGD) is efficient for the composite minimization and applicable for broad of machine learning problems. In this article, we revisit the computational complexity of this algorithm and present other novel results, especially on the convergence rates of the objective function values. The nonergodic O(1/k) rate is proved for PIGD with constant step size when the objective function is coercive. When the objective function fails to promise coercivity, we prove the sublinear rate with diminishing inertial parameters. In the case that the objective function satisfies the Polyak- Lojasiewicz (PL) property, the linear convergence is proved with much larger and general step size than the previous literature. We also extend our results to the multiblock version and present the computational complexity. Both cyclic and stochastic index selection strategies are considered.

Association of radiomic imaging features and gene expression profile as prognostic factors in pancreatic ductal adenocarcinoma.

In this study, we investigated whether radiomic features of CT image data can accurately predict HMGA2 and C-MYC gene expression status and identify the patient survival time using a machine learning approach in pancreatic ductal adenocarcinoma (PDAC). A cohort of 111 patients with PDAC was enrolled in our study. Radiomic features were extracted using conventional (shape and texture analysis) and deep learning approaches following to segmentation of preoperative CT data. To predict patient survival time, significant radiomic features were identified using a log-rank test. After surgical resection, level of HMGA2 and C-MYC gene expressions of PDAC tumor regions were classified using a support vector machines method. The model was evaluated in terms of accuracy, sensitivity, specificity, and area under the curve (AUC). Besides, inter-reader reliability analysis was used to demonstrate the robustness of the proposed features. The identified features consistently achieved good performance in survival prediction and classification of gene expression status, on images segmented by different radiologists. Using CT data from 111 patients, six features in the segmented region of images were highly correlated with survival time. Using extracted deep features of excised lesions from 47 patients, we observed an average AUC score of 0.90 with an accuracy of 95% in C-MYC prediction (sensitivity: 92% and specificity: 98%). In HGMA2 group, using shape features, the average AUC score was measured as 0.91 with an accuracy of 88% (sensitivity: 89% and specificity: 88%). In conclusion, the radiomic features of CT image can accurately predict the expression status of HMGA2 and C-MYC genes and identify the survival time of PDAC patients.

[The research on cardiac volume-time relationship based on retrospective electrocardiograph four-dimension computer tomography data collection and structured sparse algorithm].

This paper explores the relationship between the cardiac volume and time, which is applied to control dynamic heart phantom. We selected 50 patients to collect their cardiac computed tomography angiography (CTA) images, which have 20 points in time series CTA images using retrospective electrocardiograph gating, and measure the volume of four chamber in 20-time points with cardiac function analysis software. Then we grouped patients by gender, age, weight, height, heartbeat, and utilize repeated measurement design to conduct statistical analyses. We proposed structured sparse learning to estimate the mathematic expression of cardiac volume variation. The research indicates that all patients' groups are statistically significant in time factor ( P = 0.000); there are interactive effects between time and gender groups in left ventricle ( F = 8.597, P = 0.006) while no interactive effects in other chambers with the remaining groups; and the different weight groups' volume is statistically significant in right ventricle ( F = 9.004, P = 0.005) while no statistical significance in other chambers with remaining groups. The accuracy of cardiac volume and time relationship utilizing structured sparse learning is close to the least square method, but the former's expression is more concise and more robust. The number of nonzero basic function of the structured sparse model is just 2.2 percent of that of least square model. Hence, the work provides more the accurate and concise expression of the cardiac for cardiac motion simulation.

SparRec: An effective matrix completion framework of missing data imputation for GWAS.

Genome-wide association studies present computational challenges for missing data imputation, while the advances of genotype technologies are generating datasets of large sample sizes with sample sets genotyped on multiple SNP chips. We present a new framework SparRec (Sparse Recovery) for imputation, with the following properties: (1) The optimization models of SparRec, based on low-rank and low number of co-clusters of matrices, are different from current statistics methods. While our low-rank matrix completion (LRMC) model is similar to Mendel-Impute, our matrix co-clustering factorization (MCCF) model is completely new. (2) SparRec, as other matrix completion methods, is flexible to be applied to missing data imputation for large meta-analysis with different cohorts genotyped on different sets of SNPs, even when there is no reference panel. This kind of meta-analysis is very challenging for current statistics based methods. (3) SparRec has consistent performance and achieves high recovery accuracy even when the missing data rate is as high as 90%. Compared with Mendel-Impute, our low-rank based method achieves similar accuracy and efficiency, while the co-clustering based method has advantages in running time. The testing results show that SparRec has significant advantages and competitive performance over other state-of-the-art existing statistics methods including Beagle and fastPhase.