Search

This paper presents a solution to the video prediction problem based on the deep learning paradigm. Predicted frames generated by existing video prediction models are often blurry and have difficulty maintaining accuracy in multi-step prediction. To overcome these limitations, this paper presents a deep learning model, named the deep pixel restoration loop coding network (DPR-LC-Net), which employs the concept of predictive coding and adopts pixels in the real frames. While making a long-term prediction, it can generate clear prediction frames with few errors. The calculation process of DPR-LC-Net is multi-sequential: it completes calculations in the form of an approximate loop from top to bottom and then from left to right. After predicting subsequent steps, DPR-LC-Net both calculates errors and removes them from sequential prediction. Finally, the model includes a unique pixel restoration module that works efficiently on pixels in the preceding real frames to generate predicted frames, thereby improving the clarity of the predicted frames. Extensive experiments using four video datasets illustrate that the prediction performance of DPR-LC-Net is superior to that of the state-of-the-art models. [ABSTRACT FROM AUTHOR]

• We introduce a block-wise Hash difference measure to replace Euclidean distance, which significantly preserves spatial and structural information within each patch. • We design a Patch-Level Global and Local Attention-based Multi-Instance deep Learning Model that utilizes attention mechanisms to enhance Alzheimer's disease diagnosis performance and interpretability. • We construct a Patch-Level Instance Distribution-based Active Learning Strategy which aim at selecting the least discriminative samples from the candidate data set based on both sample-level and decision-level Gaussian distributions, and incorporating them into the training set to minimize labeling costs. • Experiments demonstrated the effectiveness of our proposed methods. Medical data, particularly the complex brain imaging structures, acquisition presents significant difficulties and high diagnostic expenses, resulting in a scarcity of the trainable samples in the real-world scenarios. To overcome this limitation, we present an active learning-based sampling strategy that selects the most informative samples from the unlabeled candidate sample pool for expert annotation, leading to high classification performance with a reduced number of training samples. This study adopts a patch-level perspective and introduces a multi-instance learning framework for Alzheimer's Disease diagnosis. Initially, a patch pre-selection module is designed to identify pathology-prone regions while excluding background areas and irrelevant information. Subsequently, an inner-patch local attention mechanism block and an outer-patch global attention mechanism block are developed to enhance the extraction of discriminative local and global information by the network model. Finally, an active learning sampling strategy is devised to minimize the costs associated with data acquisition and expert annotation in medical domain. The effectiveness of the proposed network framework and active learning strategy was validated through four sets of control experiments on the ADNI dataset. [ABSTRACT FROM AUTHOR]

• A novel self-supervised multi-task learning framework for medical image analysis. • An effective explicit uniformity regularizer to promote representation uniformity. • The extraction of comprehensive contextual information from anatomical structures. • Extensive experiments on chest X-ray image classification and segmentation tasks. • Visualizations of results to enhance the interpretability. Deep learning is crucial for preliminary screening and diagnostic assistance based on medical image analysis. However, limited annotated data and complex anatomical structures challenge existing models as they struggle to capture anatomical context information effectively. In response, we propose a novel self-supervised multi-task learning framework (SSMT), which integrates two key modules: a discriminative-based module and a generative-based module. These modules collaborate through multiple proxy tasks, encouraging models to learn global discriminative representations and local fine-grained representations. Additionally, we introduce an efficient uniformity regularization to further enhance the learned representations. To demonstrate the effectiveness of SSMT, we conduct extensive experiments on six public Chest X-ray image datasets. Our results highlight that SSMT not only outperforms existing state-of-the-art methods but also achieves comparable performance to the supervised model in challenging downstream tasks. The ablation study demonstrates collaboration between the key components of SSMT, showcasing its potential for advancing medical image analysis. [ABSTRACT FROM AUTHOR]

• DTr-CNN implements time series forecasting transfer learning across different datasets. • DTr-CNN alleviates the problem of lacking labeled target data in time series prediction. • Instead of only fine-tuning, DTr-CNN embeds the transfer phase into feature learning. • DTr-CNN incorporates DTW and JS divergence to evaluate similarity between datasets. • DTr-CNN takes advantages of CNN and applies it to forecasting problems. Due to the extensive practical value of time series prediction, many excellent algorithms have been proposed. Most of these methods are developed assuming that massive labeled training data are available. However, this assumption might be invalid in some actual situations. To address this limitation, a transfer learning framework with deep architectures is proposed. Since convolutional neural network (CNN) owns favorable feature extraction capability and can implement parallelization more easily, we propose a deep transfer learning method resorting to the architecture of CNN, termed as DTr-CNN for short. It can effectively alleviate the available labeled data absence and leverage useful knowledge to the current prediction. Notably, in our method, transfer learning process is implemented across different datasets. For a given target domain, in real-world scenarios, relativity of truly available potential source datasets may not be obvious, which is challenging and rarely referred to in most existing time series prediction methods. Aiming at this problem, the incorporation of Dynamic Time Warping (DTW) and Jensen-Shannon (JS) divergence is adopted for the selection of the appropriate source domain. Effectiveness of the proposed method is empirically underpinned by the experiments conducted on one group of synthetic and two groups of practical datasets. Besides, an additional experiment on NN5 dataset is conducted. [ABSTRACT FROM AUTHOR]