Abstract: With the development of plant phenomics, the identification of plant diseases from leaf images has become an effective and economic approach in plant disease science. Among the methods of plant diseases identification, the convolutional neural network (CNN) is the most popular one for its superior performance. However, CNN's representation power is still a challenge in dealing with small datasets, which greatly affects its popularization. In this work, we propose a new method, namely PiTLiD, based on pretrained Inception-V3 convolutional neural network and transfer learning to identify plant leaf diseases from phenotype data of plant leaf with small sample size. To evaluate the robustness of the proposed method, the experiments on several datasets with small-scale samples were implemented. The results show that PiTLiD performs better than compared methods. This study provides a plant disease identification tool based on a deep learning algorithm for plant phenomics. All the source data and code are accessible at https://github.com/zhanglab-wbgcas/PiTLiD .

Abstract: For the early diagnosis of hematological disorders like blood cancer, microscopic analysis of blood cells is very important. Traditional deep CNNs lead to overfitting when it receives small medical image datasets such as ALLIDB1, ALLIDB2, and ASH. This paper proposes a new and effective model for classifying and detecting Acute Lymphoblastic Leukemia (ALL) or Acute Myelogenous Leukemia (AML) that delivers excellent performance in small medical datasets. Here, we have proposed a novel Orthogonal SoftMax Layer (OSL)-based Acute Leukemia detection model that consists of ResNet 18-based deep feature extraction followed by efficient OSL-based classification. Here, OSL is integrated with the ResNet18 to improve the classification performance by making the weight vectors orthogonal to each other. Hence, it integrates ResNet benefits (residual learning and identity mapping) with the benefits of OSL-based classification (improvement of feature discrimination capability and computational efficiency). Furthermore, we have introduced extra dropout and ReLu layers in the architecture to achieve a faster network with enhanced performance. The performance verification is performed on standard ALLIDB1, ALLIDB2, and $ C\_{N}MC\_{2}019$ datasets for efficient ALL detection and ASH dataset for effective AML detection. The experimental performance demonstrates the superiority of the proposed model over other compairing models.

Abstract: Federated learning of deep neural networks has emerged as an evolving paradigm for distributed machine learning, gaining widespread attention due to its ability to update parameters without collecting raw data from users, especially in digital healthcare applications. However, the traditional centralized architecture of federated learning suffers from several problems (e.g., single point of failure, communication bottlenecks, etc.), especially malicious servers inferring gradients and causing gradient leakage. To tackle the above issues, we propose a robust and privacy-preserving decentralized deep federated learning (RPDFL) training scheme. Specifically, we design a novel ring FL structure and a Ring-Allreduce-based data sharing scheme to improve the communication efficiency in RPDFL training. Furthermore, we improve the process of distributing parameters of the Chinese residual theorem to update the execution process of the threshold secret sharing, supporting healthcare edge to drop out during the training process without causing data leakage, and ensuring the robustness of the RPDFL training under the Ring-Allreduce-based data sharing scheme. Security analysis indicates that RPDFL is provable secure. Experiment results show that RPDFL is significantly superior to standard FL methods in terms of model accuracy and convergence, and is suitable for digital healthcare applications.

Abstract: Numerous studies have reported that micro RNAs (miRNAs) play pivotal roles in disease pathogenesis based on the deregulation of the expressions of target messenger RNAs. Therefore, the identification of disease-related miRNAs is of great significance in understanding human complex diseases, which can also provide insight into the design of novel prognostic markers and disease therapies. Considering the time and cost involved in wet experiments, most recent works have focused on the effective and feasible modeling of computational frameworks to uncover miRNA-disease associations. In this study, we propose a novel framework called node2vec-based neural collaborative filtering for predicting miRNA-disease association (NCMD) based on deep neural networks. Initially, NCMD exploits Node2vec to learn low-dimensional vector representations of miRNAs and diseases. Next, it utilizes a deep learning framework that combines the linear ability of generalized matrix factorization and nonlinear ability of a multilayer perceptron. Experimental results clearly demonstrate the comparable performance of NCMD relative to the state-of-the-art methods according to statistical measures. In addition, case studies on breast cancer, lung cancer and pancreatic cancer validate the effectiveness of NCMD. Extensive experiments demonstrate the benefits of modeling a neural collaborative-filtering-based approach for discovering novel miRNA-disease associations.

Abstract: In this paper, we have presented a novel deep neural network architecture involving transfer learning approach, formed by freezing and concatenating all the layers till block4 pool layer of VGG16 pre-trained model (at the lower level) with the layers of a randomly initialized naïve Inception block module (at the higher level). Further, we have added the batch normalization, flatten, dropout and dense layers in the proposed architecture. Our transfer network, called VGGIN-Net, facilitates the transfer of domain knowledge from the larger ImageNet object dataset to the smaller imbalanced breast cancer dataset. To improve the performance of the proposed model, regularization was used in the form of dropout and data augmentation. A detailed block-wise fine tuning has been conducted on the proposed deep transfer network for images of different magnification factors. The results of extensive experiments indicate a significant improvement of classification performance after the application of fine-tuning. The proposed deep learning architecture with transfer learning and fine-tuning yields the highest accuracies in comparison to other state-of-the-art approaches for the classification of BreakHis breast cancer dataset. The articulated architecture is designed in a way that it can be effectively transfer learned on other breast cancer datasets.

Abstract: The identification of drug–target relations (DTRs) is substantial in drug development. A large number of methods treat DTRs as drug-target interactions (DTIs), a binary classification problem. The main drawback of these methods are the lack of reliable negative samples and the absence of many important aspects of DTR, including their dose dependence and quantitative affinities. With increasing number of publications of drug–protein binding affinity data recently, DTRs prediction can be viewed as a regression problem of drug–target affinities (DTAs) which reflects how tightly the drug binds to the target and can present more detailed and specific information than DTIs. The growth of affinity data enables the use of deep learning architectures, which have been shown to be among the state-of-the-art methods in binding affinity prediction. Although relatively effective, due to the black-box nature of deep learning, these models are less biologically interpretable. In this study, we proposed a deep learning-based model, named AttentionDTA, which uses attention mechanism to predict DTAs. Different from the models using 3D structures of drug–target complexes or graph representation of drugs and proteins, the novelty of our work is to use attention mechanism to focus on key subsequences which are important in drug and protein sequences when predicting its affinity. We use two separate one-dimensional Convolution Neural Networks (1D-CNNs) to extract the semantic information of drug’s SMILES string and protein’s amino acid sequence. Furthermore, a two-side multi-head attention mechanism is developed and embedded to our model to explore the relationship between drug features and protein features. We evaluate our model on three established DTA benchmark datasets, Davis, Metz, and KIBA. AttentionDTA outperforms the state-of-the-art deep learning methods under different evaluation metrics. The results show that the attention-based model can effectively extract protein features related to drug information and drug features related to protein information to better predict drug target affinities. It is worth mentioning that we test our model on IC50 dataset, which provides the binding sites between drugs and proteins, to evaluate the ability of our model to locate binding sites. Finally, we visualize the attention weight to demonstrate the biological significance of the model. The source code of AttentionDTA can be downloaded from https://github.com/zhaoqichang/AttentionDTA_TCBB .

Abstract: Heart sound analysis plays an important role in early detecting heart disease. However, manual detection requires doctors with extensive clinical experience, which increases uncertainty for the task, especially in medically underdeveloped areas. This paper proposes a robust neural network structure with an improved attention module for automatic classification of heart sound wave. In the preprocessing stage, noise removal with Butterworth bandpass filter is first adopted, and then heart sound recordings are converted into time-frequency spectrum by short-time Fourier transform (STFT). The model is driven by STFT spectrum. It automatically extracts features through four down sample blocks with different filters. Subsequently, an improved attention module based on Squeeze-and-Excitation module and coordinate attention module is developed for feature fusion. Finally, the neural network will give a category for heart sound waves based on the learned features. The global average pooling layer is adopted for reducing the model's weight and avoiding overfitting, while focal loss is further introduced as the loss function to minimize the data imbalance problem. Validation experiments have been conducted on two publicly available datasets, and the results well demonstrate the effectiveness and advantages of our method.

Abstract: Prediction of the drug–target affinity (DTA) plays an important role in drug discovery. Existing deep learning methods for DTA prediction typically leverage a single modality, namely simplified molecular input line entry specification (SMILES) or amino acid sequence to learn representations. SMILES or amino acid sequences can be encoded into different modalities. Multimodality data provide different kinds of information, with complementary roles for DTA prediction. We propose Modality-DTA, a novel deep learning method for DTA prediction that leverages the multimodality of drugs and targets. A group of backward propagation neural networks is applied to ensure the completeness of the reconstruction process from the latent feature representation to original multimodality data. The tag between the drug and target is used to reduce the noise information in the latent representation from multimodality data. Experiments on three benchmark datasets show that our Modality-DTA outperforms existing methods in all metrics. Modality-DTA reduces the mean square error by 15.7% and improves the area under the precisionrecall curve by 12.74% in the Davis dataset. We further find that the drug modality Morgan fingerprint and the target modality generated by one-hot-encoding play the most significant roles. To the best of our knowledge, Modality-DTA is the first method to explore multimodality for DTA prediction.

Abstract: Enhancer, a distal cis-regulatory element controls gene expression. Experimental prediction of enhancer elements is time-consuming and expensive. Consequently, various inexpensive deep learning-based fast methods have been developed for predicting the enhancers and determining their strength. In this paper, we have proposed a two-stage deep learning-based framework leveraging DNA structural features, natural language processing, convolutional neural network, and long short-term memory to predict the enhancer elements accurately in the genomics data. In the first stage, we extracted the features from DNA sequence data by using three feature representation techniques viz., k-mer based feature extraction along with word2vector based interpretation of underlined patterns, one-hot encoding, and the DNAshape technique. In the second stage, strength of enhancers is predicted from the extracted features using a hybrid deep learning model. The method is capable of adapting itself to varying sizes of datasets. Also, as proposed model can capture long-range sequencing patterns, the robustness of the method remains unaffected against minor variations in the genomics sequence. The method outperforms the other state-of-the-art methods at both stages in terms of performance metrics of prediction accuracy, specificity, Mathew's correlation coefficient, and area under the ROC curve. In summary, the proposed method is a reliable method for enhancer prediction.

Abstract: Deep learning methods have achieved great success in medical image analysis domain. However, most of them suffer from slow convergency and high computing cost, which prevents their further widely usage in practical scenarios. Moreover, it has been proved that exploring and embedding context knowledge in deep network can significantly improve accuracy. To emphasize these tips, we present CDT-CAD, i.e., context-aware deformable transformers for end-to-end chest abnormality detection on X-Ray images. CDT-CAD firstly constructs an iterative context-aware feature extractor, which not only enlarges receptive fields to encode multi-scale context information via dilated context encoding blocks, but also captures unique and scalable feature variation patterns in wavelet frequency domain via frequency pooling blocks. Afterwards, a deformable transformer detector on the extracted context features is built to accurately classify disease categories and locate regions, where a small set of key points are sampled, thus leading the detector to focus on informative feature subspace and accelerate convergence speed. Through comparative experiments on Vinbig Chest and Chest Det 10 Datasets, CDT-CAD demonstrates its effectiveness in recognizing chest abnormities and outperforms 1.4% and 6.0% than the existing methods in AP50 and AR on VinBig dateset, and 0.9% and 2.1% on Chest Det-10 dataset, respectively.

Abstract: Previous studies have confirmed microRNA (miRNA), small single-stranded non-coding RNA, participates in various biological processes and plays vital roles in many complex human diseases. Therefore, developing an efficient method to infer potential miRNA disease associations could greatly help understand operational mechanisms for diseases at the molecular level. However, during these early stages for miRNA disease prediction, traditional biological experiments are laborious and expensive. Therefore, this study proposes a novel method called AGAEMD (node-level Attention Graph Auto-Encoder to predict potential MiRNA Disease associations). We first create a heterogeneous matrix incorporating miRNA similarity, disease similarity, and known miRNA-disease associations. Then these matrixes are input into a node-level attention encoder-decoder network which utilizes low dimensional dense embeddings to represent nodes and calculate association scores. To verify the effectiveness of the proposed method, we conduct a series of experiments on two benchmark datasets (the Human MicroRNA Disease Database v2.0 and v3.2) and report the averages over 10 runs in comparison with several state-of-the-art methods. Experimental results have demonstrated the excellent performance of AGAEMD in comparison with other methods. Three important diseases (Colon Neoplasms, Lung Neoplasms, Lupus Vulgaris) were applied in case studies. The results comfirm the reliable predictive performance of AGAEMD.

Abstract: N6-methyladenosine (m6A) is a common post-transcriptional alteration that plays a critical function in a variety of biological processes. Although experimental approaches for identifying m6A sites have been developed and deployed, they are currently expensive for transcriptome-wide m6A identification. Some computational strategies for identifying m6A sites have been presented as an effective complement to the experimental procedure. However, their performance still requires improvement. In this study, we have proposed a novel tool called DL-m6A for the identification of m6A sites in mammals using deep learning based on different encoding schemes. The proposed tool uses three encoding schemes which give the required contextual feature representation to the input RNA sequence. Later these contextual feature vectors individually go through several neural network layers for shallow feature extraction after which they are concatenated to a single feature vector. The concatenated feature map is then used by several other layers to extract the deep features so that the insight features of the sequence can be used for the prediction of m6A sites. The proposed tool is firstly evaluated on the tissue-specific dataset and later on a full transcript dataset. To ensure the generalizability of the tool we assessed the proposed model by training it on a full transcript dataset and test on the tissue-specific dataset. The achieved results by the proposed model have outperformed the existing tools. The results demonstrate that the proposed tool can be of great use for the biology experts and therefore a freely accessible web-server is created which can be accessed at: http://nsclbio.jbnu.ac.kr/tools/DL-m6A/ .

Abstract: Membrane proteins are the main undertaker of biomembrane functions and play a vital role in many biological activities of organisms. Prediction of membrane protein types has a great help in determining the function of proteins and understanding the interactions of membrane proteins. However, the biochemical experiment is expensive and not suitable for the large-scale identification of membrane protein types. Therefore, computational methods were used to improve the efficiency of biological experiments. Most existing computational methods only use a single feature of protein, or use multiple features but do not integrate these well. In our study, the protein sequence is described via three different views (features), including amino acid composition, evolutionary information and physicochemical properties of amino acids. To exploit information among all views (features), we introduce a coupling strategy for Kernel Sparse Representation based Classification (KSRC) and construct a new model called Multi-view KSRC (MvKSRC). We implement our method on 4 benchmark data sets of membrane proteins. The comparison results indicate that our method is much superior to all existing methods.

Abstract: Protein-protein interactions (PPIs) play a critical role in the proteomics study, and a variety of computational algorithms have been developed to predict PPIs. Though effective, their performance is constrained by high false-positive and false-negative rates observed in PPI data. To overcome this problem, a novel PPI prediction algorithm, namely PASNVGA, is proposed in this work by combining the sequence and network information of proteins via variational graph autoencoder. To do so, PASNVGA first applies different strategies to extract the features of proteins from their sequence and network information, and obtains a more compact form of these features using principal component analysis. In addition, PASNVGA designs a scoring function to measure the higher-order connectivity between proteins and so as to obtain a higher-order adjacency matrix. With all these features and adjacency matrices, PASNVGA trains a variational graph autoencoder model to further learn the integrated embeddings of proteins. The prediction task is then completed by using a simple feedforward neural network. Extensive experiments have been conducted on five PPI datasets collected from different species. Compared with several state-of-the-art algorithms, PASNVGA has been demonstrated as a promising PPI prediction algorithm. The source code of PASNVGA and all datasets are available at https://github.com/weizhi-code/PASNVGA.

Abstract: Therapeutic peptide prediction is critical for drug development and therapeutic therapy. Researchers have developed several computational methods to identify different therapeutic peptide types. However, most computational methods focus on identifying the specific type of therapeutic peptides and fail to accurately predict all types of therapeutic peptides. Moreover, it is still challenging to utilize different properties features to predict the therapeutic peptides. In this study, a novel stacking framework PreTP-Stack is proposed for predicting different types of therapeutic peptide. PreTP-Stack is constructed based on ten different features and four predictors (Random Forest, Linear Discriminant Analysis, XGBoost and Support Vector Machine). Then the proposed method constructs an auto-weighted multi-view learning model as a final meta-classifier to enhance the performance of the basic models. Experimental results showed that the proposed method achieved better or highly comparable performance with the state-of-the-art methods for predicting eight types of therapeutic peptides A user-friendly web-server predictor is available at http://bliulab.net/PreTP-Stack.

Abstract: Apple leaf diseases seriously affect the quality of apples and may lead to yield losses, detecting apple leaf diseases accurately can prevent diseases from spreading and promote the healthy growth of the industry. However, recent studies cannot achieve accurate detection of leaf diseases with high accuracy because the lesions are of different sizes. So, this paper proposed a novel apple leaf disease detection method called VMF-SSD (V-space-based Multi-scale Feature-fusion SSD), which is designed to extract more reliable multi-scale feature representations for varied sizes of diseased spots and improve the final detection performance. The multi-scale feature extraction is established with multi-scale feature representation to further improve the disease detection performance, especially for small spots. After that, a V-space-based location branch is presented to enhance the texture feature information and help further identify disease spot location. Finally, attention mechanisms are utilized to automatically learn the importance of feature channels at different scales for distinguishing diseased spots of different sizes. Experimental results showed that the VMF-SSD method achieves 83.19% mAP and obtains the detection speed of 27.53 FPS on the test set, which indicates that the proposed VMF-SSD method can achieve competitive performance on apple leaf diseases detection task and satisfy the requirements of agricultural production applications.

Abstract: Biological processes are often modelled using ordinary differential equations. The unknown parameters of these models are estimated by optimizing the fit of model simulation and experimental data. The resulting parameter estimates inevitably possess some degree of uncertainty. In practical applications it is important to quantify these parameter uncertainties as well as the resulting prediction uncertainty, which are uncertainties of potentially time-dependent model characteristics. Unfortunately, estimating prediction uncertainties accurately is nontrivial, due to the nonlinear dependence of model characteristics on parameters. While a number of numerical approaches have been proposed for this task, their strengths and weaknesses have not been systematically assessed yet. To fill this knowledge gap, we apply four state of the art methods for uncertainty quantification to four case studies of different computational complexities. This reveals the trade-offs between their applicability and their statistical interpretability. Our results provide guidelines for choosing the most appropriate technique for a given problem and applying it successfully.

Abstract: Multi-layer networks provide an effective and efficient tool to model and characterize complex systems with multiple types of interactions, which differ greatly from the traditional single-layer networks. Graph clustering in multi-layer networks is highly non-trivial since it is difficult to balance the connectivity of clusters and the connection of various layers. The current algorithms for the layer-specific clusters are criticized for the low accuracy and sensitivity to the perturbation of networks. To overcome these issues, a novel algorithm for the layer-specific module in multi-layer networks based on nonnegative matrix factorization (LSNMF) is proposed by explicitly exploring the specific features of vertices. LSNMF first extract features of vertices in multi-layer networks by using nonnegative matrix factorization (NMF) and then decompose features of vertices into the common and specific components. The orthogonality constraint is imposed on the specific components to ensure the specificity of features of vertices, which provides a better strategy to characterize and model the structure of layer-specific modules. The extensive experiments demonstrate that the proposed algorithm dramatically outperforms state-of-the-art baselines in terms of various measurements. Furthermore, LSNMF efficiently extracts stage-specific modules, which are more likely to enrich the known functions, and also associate with the survival time of patients.

Abstract: The ability to identify and characterize not only the protein-protein interactions but also their internal modular organization through network analysis is fundamental for understanding the mechanisms of biological processes at the molecular level. Indeed, the detection of the network communities can enhance our understanding of the molecular basis of disease pathology, and promote drug discovery and disease treatment in personalized medicine. This work gives an overview of recent computational methods for the detection of protein complexes and functional modules in protein-protein interaction networks, also providing a focus on some of its applications. We propose a systematic reformulation of frequently adopted taxonomies for these methods, also proposing new categories to keep up with the most recent research. We review the literature of the last five years (2017-2021) and provide links to existing data and software resources. Finally, we survey recent works exploiting module identification and analysis, in the context of a variety of disease processes for biomarker identification and therapeutic target detection. Our review provides the interested reader with an up-to-date and self-contained view of the existing research, with links to state-of-the-art literature and resources, as well as hints on open issues and future research directions in complex detection and its applications.

Abstract: Depression is a mental disorder characterized by persistent depressed mood or loss of interest in performing activities, causing significant impairment in daily routine. Possible causes include psychological, biological, and social sources of distress. Clinical depression is the more-severe form of depression, also known as major depression or major depressive disorder. Recently, electroencephalography and speech signals have been used for early diagnosis of depression; however, they focus on moderate or severe depression. We have combined audio spectrogram and multiple frequencies of EEG signals to improve diagnostic performance. To do so, we have fused different levels of speech and EEG features to generate descriptive features and applied vision transformers and various pre-trained networks on the speech and EEG spectrum. We have conducted extensive experiments on Multimodal Open Dataset for Mental-disorder Analysis (MODMA) dataset, which showed significant improvement in performance in depression diagnosis (0.972, 0.973 and 0.973 precision, recall and F1 score respectively ) for patients at the mild stage. Besides, we provided a web-based framework using Flask and provided the source code publicly https://github.com/RespectKnowledge/EEG_Speech_Depression_MultiDL.

Abstract: Detecting potential associations between drugs and diseases plays an indispensable role in drug development, which has also become a research hotspot in recent years. Compared with traditional methods, some computational approaches have the advantages of fast speed and low cost, which greatly accelerate the progress of predicting the drug-disease association. In this study, we propose a novel similarity-based method of low-rank matrix decomposition based on multi-graph regularization. On the basis of low-rank matrix factorization with L2 regularization, the multi-graph regularization constraint is constructed by combining a variety of similarity matrices from drugs and diseases respectively. In the experiments, we analyze the difference in the combination of different similarities, resulting that combining all the similarity information on drug space is unnecessary, and only a part of the similarity information can achieve the desired performance. Then our method is compared with other existing models on three data sets (Fdataset, Cdataset and LRSSLdataset) and have a good advantage in the evaluation measurement of AUPR. Besides, a case study experiment is conducted and showing that the superior ability for predicting the potential disease-related drugs of our model. Finally, we compare our model with some methods on six real world datasets, and our model has a good performance in detecting real world data.

Abstract: As the segment of diseased tissue in PET images is time-consuming, laborious and low accuracy, this work proposes an automated framework for PET image screening, denoising and diseased tissue segmentation. First, taking into account the characteristics of PET images, the framework uses a differential activation filter to select whole-body images containing lesion tissue. Second, a new neural network containing residual connections which has powerful generalization performance compared with normal FCN network is proposed for PET image reconstruction and denoising. Finally, in the segmentation of lesion tissues, a custom clustering algorithm based on the density is used to distinguishe the lesion tissue part from the normal tissue. Tests on real medical PET images show that the whole automated framework has good performance and time cost in PET lesion image screening, image denoising and lesion tissue segmentation compared with other algorithms. The framework shows promising scientific study and application prospects.

Abstract: A growing number of studies show that the human microbiome plays a vital role in human health and can be a crucial factor in predicting certain human diseases. However, microbiome data are often characterized by the limited samples and high-dimensional features, which pose a great challenge for machine learning methods. Therefore, this paper proposes a novel ensemble deep learning disease prediction method that combines unsupervised and supervised learning paradigms. First, unsupervised deep learning methods are used to learn the potential representation of the sample. Afterwards, the disease scoring strategy is developed based on the deep representations as the informative features for ensemble analysis. To ensure the optimal ensemble, a score selection mechanism is constructed, and performance boosting features are engaged with the original sample. Finally, the composite features are trained with gradient boosting classifier for health status decision. For case study, the ensemble deep learning flowchart has been demonstrated on six public datasets extracted from the human microbiome profiling. The results show that compared with the existing algorithms, our framework achieves better performance on disease prediction.

Abstract: Achieving accurate retinal vessel segmentation is critical in the progression and diagnosis of vision-threatening diseases such as diabetic retinopathy and age-related macular degeneration. Existing vessel segmentation methods are based on encoder-decoder architectures, which frequently fail to take into account the retinal vessel structure's context in their analysis. As a result, such methods have difficulty bridging the semantic gap between encoder and decoder characteristics. This paper proposes a Prompt Deep Light-weight Vessel Segmentation Network (PLVS-Net) to address these issues by using prompt blocks. Each prompt block use combination of asymmetric kernel convolutions, depth-wise separable convolutions, and ordinary convolutions to extract useful features. This novel strategy improves the performance of the segmentation network while simultaneously decreasing the number of trainable parameters. Our method outperformed competing approaches in the literature on three benchmark datasets, including DRIVE, STARE, and CHASE.

Abstract: Automated recognition of Human Phenotype Ontology (HPO) terms from clinical texts is of significant interest to the field of clinical data mining. In this study, we develop a combined deep learning method named PhenoBERT for this purpose. PhenoBERT uses BERT, currently the state-of-the-art NLP model, as its core model for evaluating whether a clinically relevant text segment (CTS) could be represented by an HPO term. However, to avoid unnecessary comparison of a CTS with each of ∼14,000 HPO terms using BERT, we introduce a two-levels CNN module consisting of a series of CNN models organized at two levels in PhenoBERT. For a given CTS, the CNN module produces only a short list of candidate HPO terms for BERT to evaluate, significantly improving the computational efficiency. In addition, BERT is able to assign an ancestor HPO term to a CTS when recognition of the direct HPO term is not successful, mimicking the process of HPO term assignment by human. In two benchmarks, PhenoBERT outperforms four traditional dictionary-based methods and two recently developed deep learning-based methods in two benchmark tests, and its advantage is more obvious when the recognition task is more challenging. As such, PhenoBERT is of great use for assisting in the mining of clinical text data.

Abstract: Analysis of neuroimaging data (e.g., Magnetic Resonance Imaging, structural and functional MRI) plays an important role in monitoring brain dynamics and probing brain structures. Neuroimaging data are multi-featured and non-linear by nature, and it is a natural way to organise these data as tensors prior to performing automated analyses such as discrimination of neurological disorders like Parkinson's Disease (PD) and Attention Deficit and Hyperactivity Disorder (ADHD). However, the existing approaches are often subject to performance bottlenecks (e.g., conventional feature extraction and deep learning based feature construction), as these can lose the structural information that correlates multiple data dimensions or/and demands excessive empirical and application-specific settings. This study proposes a Deep Factor Learning model on a Hilbert Basis tensor (namely, HB-DFL) to automatically derive latent low-dimensional and concise factors of tensors. This is achieved through the application of multiple Convolutional Neural Networks (CNNs) in a non-linear manner along all possible dimensions with no assumed a priori knowledge. HB-DFL leverages the Hilbert basis tensor to enhance the stability of the solution by regularizing the core tensor to allow any component in a certain domain to interact with any component in the other dimensions. The final multi-domain features are handled through another multi-branch CNN to achieve reliable classification, exemplified here using MRI discrimination as a typical case. A case study of MRI discrimination has been performed on public MRI datasets for discrimination of PD and ADHD. Results indicate that 1) HB-DFL outperforms the counterparts in terms of FIT, mSIR and stability (mSC and umSC) of factor learning; 2) HB-DFL identifies PD and ADHD with an accuracy significantly higher than state-of-the-art methods do. Overall, HB-DFL has significant potentials for neuroimaging data analysis applications with its stability of automatic construction of structural features.

Abstract: Lung cancer is the leading cause of cancer death worldwide. The best solution for lung cancer is to diagnose the pulmonary nodules in the early stage, which is usually accomplished with the aid of thoracic computed tomography (CT). As deep learning thrives, convolutional neural networks (CNNs) have been introduced into pulmonary nodule detection to help doctors in this labor-intensive task and demonstrated to be very effective. However, the current pulmonary nodule detection methods are usually domain-specific, and cannot satisfy the requirement of working in diverse real-world scenarios. To address this issue, we propose a slice grouped domain attention (SGDA) module to enhance the generalization capability of the pulmonary nodule detection networks. This attention module works in the axial, coronal, and sagittal directions. In each direction, we divide the input feature into groups, and for each group, we utilize a universal adapter bank to capture the feature subspaces of the domains spanned by all pulmonary nodule datasets. Then the bank outputs are combined from the perspective of domain to modulate the input group. Extensive experiments demonstrate that SGDA enables substantially better multi-domain pulmonary nodule detection performance compared with the state-of-the-art multi-domain learning methods.