20 August 2020 in European Heart Journal
Aims: Facial features were associated with increased risk of coronary artery disease (CAD). We developed and validated a deep learning algorithm for detecting CAD based on facial photos. Methods and results: We conducted a multicentre cross-sectional study of patients undergoing coronary angiography or computed tomography angiography at nine Chinese sites to train and validate a deep convolutional neural network for the detection of CAD (at least one ≥50% stenosis) from patient facial photos. Between July 2017 and March 2019, 5796 patients from eight sites were consecutively enrolled and randomly divided into training (90%, n = 5216) and validation (10%, n = 580) groups for algorithm development. Between April 2019 and July 2019, 1013 patients from nine sites were enrolled in test group for algorithm test. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated using radiologist diagnosis as the reference standard. Using an operating cut point with high sensitivity, the CAD detection algorithm had sensitivity of 0.80 and specificity of 0.54 in the test group; the AUC was 0.730 (95% confidence interval, 0.699–0.761). The AUC for the algorithm was higher than that for the Diamond–Forrester model (0.730 vs. 0.623, P < 0.001) and the CAD consortium clinical score (0.730 vs. 0.652, P < 0.001). Conclusion: Our results suggested that a deep learning algorithm based on facial photos can assist in CAD detection in this Chinese cohort. This technique may hold promise for pre-test CAD probability assessment in outpatient clinics or CAD screening in community. Further studies to develop a clinical available tool are warranted.
21 March 2025 in IEEE Transactions on Pattern Analysis and Machine Intelligence
6D pose estimation of rigid objects is a long-standing and challenging task in computer vision. Recently, the emergence of deep learning reveals the potential of Convolutional Neural Networks (CNNs) to predict reliable 6D poses. Given that direct pose regression networks currently exhibit suboptimal performance, most methods still resort to traditional techniques to varying degrees. For example, top-performing methods often adopt an indirect strategy by first establishing 2D-3D or 3D-3D correspondences followed by applying the RANSAC-based P $n$ P or Kabsch algorithms, and further employing ICP for refinement. Despite the performance enhancement, the integration of traditional techniques makes the networks time-consuming and not end-to-end trainable. Orthogonal to them, this paper introduces a fully learning-based object pose estimator. In this work, we first perform an in-depth investigation of both direct and indirect methods and propose a simple yet effective Geometry-guided Direct Regression Network (GDRN) to learn the 6D pose from monocular images in an end-to-end manner. Afterwards, we introduce a geometry-guided pose refinement module, enhancing pose accuracy when extra depth data is available. Guided by the predicted coordinate map, we build an end-to-end differentiable architecture that establishes robust and accurate 3D-3D correspondences between the observed and rendered RGB-D images to refine the pose. Our enhanced pose estimation pipeline GDRNPP (GDRN Plus Plus) conquered the leaderboard of the BOP Challenge for two consecutive years, becoming the first to surpass all prior methods that relied on traditional techniques in both accuracy and speed.
17 June 2019 in IEEE Access
Although classification-based approaches achieved remarkable performance in viewpoints estimation, they still have limitations and face the challenging viewpoints ambiguity problem. In this paper, we analyze this problem and give solutions: 1) we propose Viewpoint Discernibility Matrix (VDM) loss, which is a more suitable loss than the one-hot cross-entropy loss by tolerating the sub-optimal predictions and penalizing the wrong predictions on ambiguous viewpoints; and 2) we propose Auxiliary Hierarchical Viewpoints Supervision (AHVS) method, which is able to restrain the network to pay closer attention to the features of ambiguous viewpoints. Training with VDM loss and AHVS, the model is endowed with strong representation ability to achieve significant improvements on viewpoints estimation. Extensive experiments are conducted and show the superiority of our approach. On the large Pascal3D+ dataset, we achieve the state-of-the-art results among all commonly used metrics.
22 November 2025 in Pattern Recognition