Computer-aided technologies have achieved excellent results in lung diseases screening. However, most of the existing studies models the known type of disease, which can lead to the risk of misdiagnosis and missed diagnosis to the unknown type of disease. In addition, current research primarily focuses on achieving high accuracy without solving the problems of misdiagnosis and missed diagnosis, making it difficult to apply in clinical practice. This paper proposes a more clinically applicable goal for computer-aided pulmonary disease screening by combing the doctor’s diagnosis process: to ensure a zero missed diagnosis rate while reducing the misdiagnosis rate. To achieve the screening goal mentioned above, this paper proposes an improved variational autoencoder network based on one-class classification for preliminary screening of lung diseases and extraction of deep encoding features from X-ray images. Then, by combining the complementary advantages between radiomics features based on doctors’ experience and deep learning features, an ensemble learning model is constructed to complete the final lung disease screening. The method in this study improved the classification performance and reduced the missed diagnosis rate with only normal X-ray images involved in training. Experimental results show that the proposed method achieves an AUC value of 0.9848±0.0023 and reduces the misdiagnosis rate to 0.1498±0.0057 when the missed diagnosis rate is zero, further validating that the method can effectively achieve the pulmonary diseases screening goal of this paper. At the same time, we compared the screening results of the ensemble model and individual deep learning model. We found that the results of the ensemble model was significantly better than the deep learning model, which further highlights the effectiveness of integrating doctors’ experience.

Due to the increasing number and complexity of diagnostic imaging studies, radiologists are now faced with a heavy workload in their daily image reviewing task. It has become a major concern to reduce the burden of image reviewing process for radiologists and help them improve the efficiency of diagnosis. The recognition and understanding of interactive intent during radiologists image review can be used to develop new intelligent review systems that can provide adaptive tools and optimize user workflow to help radiologists perform digital image review tasks more efficiently and accurately. Currently, radiologists still primarily use a graphical user interface based on mouse input for medical image reviewing. To this end, this paper proposes an intention understanding method of medical image reviewing based on mouse interaction trajectories. Firstly, semantic segmentation of mouse interaction trajectories is performed, and then the semantic annotation of mouse interaction behaviors is implemented based on logical reasoning and supervised learning for the segmented trajectory subsequences. The method is validated on a self-built mouse trajectory dataset under the TIoU threshold of [0.3∶0.8∶0.1], and the average recall, precision and F-score of the method reach 86.7%, 83.1% and 85.4%, respectively. The results show that the method can effectively identify the interactive intention based on the implicit mouse interaction data, which confirms the feasibility and effectiveness of the interactive intention recognition method based on semantic understanding of mouse interaction trajectories in the field of medical images assisted reviewing research.