BACKGROUND:Given the biases and ethical concerns of AI models, the fully automatic diagnosis of diseases in clinical settings is challenging. In contrast, clinician-AI collaboration is considered essential to ensure the validity and reliability of utilizing AI models in clinical practice. However, effective strategies for clinician-AI collaboration remain largely unexplored. METHODS:This study proposed a three-step general clinician-AI collaboration pipeline aimed at improving disease diagnosis efficiency: first, utilizing large real-world clinical datasets to evaluate and clarify clinicians' diagnostic strengths/weaknesses; second, developing an AI model to complement clinicians' weakness in disease diagnosis; and finally, proposing a clinician-AI collaboration strategy to leverage the strengths of both AI and clinicians. The effectiveness of this pipeline was validated through a study focusing on clinical paroxysmal atrial fibrillation (PAF) detection, utilizing 24-h Holter recordings from over 30,000 patients. FINDINGS:In PAF detection, clinicians alone required a significant amount of time to identify the data and still overlooked 13.7% of PAF patients but successfully identified all non-atrial fibrillation (AF) patients. Conversely, AI alone rarely missed PAF patients but misidentified 23.3% of non-AF patients as having PAF. After implementing the proposed clinician-AI collaboration strategy, all patients were correctly identified, and clinicians' workload was reduced by 76.7%. CONCLUSIONS:This study improves both the efficiency and reliability of PAF detection, bridging the gap between AI model development and its clinical application, thereby effectively promoting the application of AI models in clinical AF screening. FUNDING:This study was supported in part by the National Natural Science Foundation of China.

BACKGROUND:Early diagnosis of atrial fibrillation (AF) is important for preventing stroke and other complications. Predicting AF risk in advance can improve early diagnostic efficiency. Deep learning has been used for disease risk prediction; however, it lacks adherence to evidence-based medicine standards. Identifying the underlying mechanisms behind disease risk prediction is important and required. METHODS:We developed an explainable deep learning model called HBBI-AI to predict AF risk using only heart beat-to-beat intervals (HBBIs) during sinus rhythm. We proposed a possible AF mechanism based on the model's explainability and verified this conjecture using confirmed AF risk factors while also examining new AF risk factors. Finally, we investigated the changes in clinicians' ability to predict AF risk using only HBBIs before and after learning the model's explainability. FINDINGS:HBBI-AI consistently performed well across large in-house and external public datasets. HBBIs with large changes or extreme stability were critical predictors for increased AF risk, and the underlying cause was autonomic imbalance. We verified various AF risk factors and discovered that autonomic imbalance was associated with all these factors. Finally, cardiologists effectively understood and learned from these findings to improve their abilities in AF risk prediction. CONCLUSIONS:HBBI-AI effectively predicted AF risk using only HBBI information through evaluating autonomic imbalance. Autonomic imbalance may play an important role in many risk factors of AF rather than in a limited number of risk factors. FUNDING:This study was supported in part by the National Key R&D Program and the National Natural Science Foundation of China.

Semi-supervised learning via teacher-student network can train a model effectively on a few labeled samples. It enables a student model to distill knowledge from the teacher's predictions of extra unlabeled data. However, such knowledge flow is typically unidirectional, having the accuracy vulnerable to the quality of teacher model. In this paper, we seek to robust 3D reconstruction of stereo endoscopic images by proposing a novel fashion of bidirectional learning between two learners, each of which can play both roles of teacher and student concurrently. Specifically, we introduce two self-supervisions, i.e., Adaptive Cross Supervision (ACS) and Adaptive Parallel Supervision (APS), to learn a dual-branch convolutional neural network. The two branches predict two different disparity probability distributions for the same position, and output their expectations as disparity values. The learned knowledge flows across branches along two directions: a cross direction (disparity guides distribution in ACS) and a parallel direction (disparity guides disparity in APS). Moreover, each branch also learns confidences to dynamically refine its provided supervisions. In ACS, the predicted disparity is softened into a unimodal distribution, and the lower the confidence, the smoother the distribution. In APS, the incorrect predictions are suppressed by lowering the weights of those with low confidence. With the adaptive bidirectional learning, the two branches enjoy well-tuned mutual supervisions, and eventually converge on a consistent and more accurate disparity estimation. The experimental results on four public datasets demonstrate our superior accuracy over other state-of-the-arts with a relative decrease of averaged disparity error by at least 9.76%.

In recent years, deep learning (DL)-based registration technology has significantly improved the calculation speed of medical image registration. Existing DL-based registration methods generally use raw data features to predict the deformation field. However, this strategy may not be very effective for difficult registration tasks. Hence, in this study, we propose a similarity attention-based convolutional neural network (CNN) for accurate and robust three-dimensional medical image registration. We first introduce a similarity-based local attention model as an auxiliary module for building a displacement searching space, instead of a direct displacement prediction based on raw data. The proposed model can help the network focus on spatial correspondences with high similarities and ignore those with low similarities. A multi-scale CNN is then integrated with the similarity-based local attention for providing non-local attention, lightweight network, and coarse-to-fine registration. We evaluated the proposed method for various applications, such as the registration of large-scope abdominal computerized tomography (CT) images and chest CT images acquired at different respiratory phases, and atlas registration in magnetic resonance imaging. The experimental results demonstrate that the proposed method can provide a more accurate and robust registration performance than state-of-the-art registration methods.

Automated Cobb angle estimation on X-ray images is crucial to scoliosis diagnosis. The existing efforts are typically two extremes, which either laboriously detect the raw vertebral landmarks or directly regress Cobb angles from the entire image. In this paper, we propose a novel two-stage end-to-end method as a balanced solution, to avoid vulnerability to false landmarks, and to preserve flexibility in clinical usages. Concretely, we cascade two stages sequentially for detecting vertebrae and then regressing their bending directions instead of raw landmarks. In the detection stage, we combine two networks called LocNet and SegNet to robustly localize vertebrae, and meanwhile to suppress the false positives by additionally segmenting the whole spine. In the subsequent stage, we introduce a regression network named RegNet to accurately regress bending directions of localized vertebrae. Furthermore, the vertebra-aligned local regions on LocNet's intermediate features are cropped via RoIAlign-pooling, and RegNet inherits the cropped regions to learn only feature residuals. By doing so, the regression difficulty can be dramatically alleviated, and the two stages are deeply coupled and mutually guided in an end-to-end training. Moreover, a random perturbation on the inherited features further enhances RegNet's robustness. We benchmark our method on both public and private datasets, and the errors are 2.92 $\pm$ 2.34 $^{\circ }$ and 6.87 $\pm$ 6.26% in terms of CMAE and SMAPE on the widely-employed AASCE dataset, outperforming other state-of-the-arts by at least 16.81% and 6.15%, respectively. Also, a clinical user study verifies our promising flexibility for allowing convenient rectifications to further decrease errors by a large marge.