A survey on the algorithm and development of multiple sequence alignment.

Multiple sequence alignment (MSA) is an essential cornerstone in bioinformatics, which can reveal the potential information in biological sequences, such as function, evolution and structure. MSA is widely used in many bioinformatics scenarios, such as phylogenetic analysis, protein analysis and genomic analysis. However, MSA faces new challenges with the gradual increase in sequence scale and the increasing demand for alignment accuracy. Therefore, developing an efficient and accurate strategy for MSA has become one of the research hotspots in bioinformatics. In this work, we mainly summarize the algorithms for MSA and its applications in bioinformatics. To provide a structured and clear perspective, we systematically introduce MSA's knowledge, including background, database, metric and benchmark. Besides, we list the most common applications of MSA in the field of bioinformatics, including database searching, phylogenetic analysis, genomic analysis, metagenomic analysis and protein analysis. Furthermore, we categorize and analyze classical and state-of-the-art algorithms, divided into progressive alignment, iterative algorithm, heuristics, machine learning and divide-and-conquer. Moreover, we also discuss the challenges and opportunities of MSA in bioinformatics. Our work provides a comprehensive survey of MSA applications and their relevant algorithms. It could bring valuable insights for researchers to contribute their knowledge to MSA and relevant studies.

A novel convolution attention model for predicting transcription factor binding sites by combination of sequence and shape.

The discovery of putative transcription factor binding sites (TFBSs) is important for understanding the underlying binding mechanism and cellular functions. Recently, many computational methods have been proposed to jointly account for DNA sequence and shape properties in TFBSs prediction. However, these methods fail to fully utilize the latent features derived from both sequence and shape profiles and have limitation in interpretability and knowledge discovery. To this end, we present a novel Deep Convolution Attention network combining Sequence and Shape, dubbed as D-SSCA, for precisely predicting putative TFBSs. Experiments conducted on 165 ENCODE ChIP-seq datasets reveal that D-SSCA significantly outperforms several state-of-the-art methods in predicting TFBSs, and justify the utility of channel attention module for feature refinements. Besides, the thorough analysis about the contribution of five shapes to TFBSs prediction demonstrates that shape features can improve the predictive power for transcription factors-DNA binding. Furthermore, D-SSCA can realize the cross-cell line prediction of TFBSs, indicating the occupancy of common interplay patterns concerning both sequence and shape across various cell lines. The source code of D-SSCA can be found at https://github.com/MoonLord0525/.

HAMPLE: deciphering TF-DNA binding mechanism in different cellular environments by characterizing higher-order nucleotide dependency.

MOTIVATION: Transcription factor (TF) binds to conservative DNA binding sites in different cellular environments and development stages by physical interaction with interdependent nucleotides. However, systematic computational characterization of the relationship between higher-order nucleotide dependency and TF-DNA binding mechanism in diverse cell types remains challenging. RESULTS: Here, we propose a novel multi-task learning framework HAMPLE to simultaneously predict TF binding sites (TFBS) in distinct cell types by characterizing higher-order nucleotide dependencies. Specifically, HAMPLE first represents a DNA sequence through three higher-order nucleotide dependencies, including k-mer encoding, DNA shape and histone modification. Then, HAMPLE uses the customized gate control and the channel attention convolutional architecture to further capture cell-type-specific and cell-type-shared DNA binding motifs and epigenomic languages. Finally, HAMPLE exploits the joint loss function to optimize the TFBS prediction for different cell types in an end-to-end manner. Extensive experimental results on seven datasets demonstrate that HAMPLE significantly outperforms the state-of-the-art approaches in terms of auROC. In addition, feature importance analysis illustrates that k-mer encoding, DNA shape, and histone modification have predictive power for TF-DNA binding in different cellular environments and are complementary to each other. Furthermore, ablation study, and interpretable analysis validate the effectiveness of the customized gate control and the channel attention convolutional architecture in characterizing higher-order nucleotide dependencies. AVAILABILITY AND IMPLEMENTATION: The source code is available at https://github.com/ZhangLab312/Hample.

High-resolution transcription factor binding sites prediction improved performance and interpretability by deep learning method.

Transcription factors (TFs) are essential proteins in regulating the spatiotemporal expression of genes. It is crucial to infer the potential transcription factor binding sites (TFBSs) with high resolution to promote biology and realize precision medicine. Recently, deep learning-based models have shown exemplary performance in the prediction of TFBSs at the base-pair level. However, the previous models fail to integrate nucleotide position information and semantic information without noisy responses. Thus, there is still room for improvement. Moreover, both the inner mechanism and prediction results of these models are challenging to interpret. To this end, the Deep Attentive Encoder-Decoder Neural Network (D-AEDNet) is developed to identify the location of TFs-DNA binding sites in DNA sequences. In particular, our model adopts Skip Architecture to leverage the nucleotide position information in the encoder and removes noisy responses in the information fusion process by Attention Gate. Simultaneously, the Transcription Factor Motif Discovery based on Sliding Window (TF-MoDSW), an approach to discover TFs-DNA binding motifs by utilizing the output of neural networks, is proposed to understand the biological meaning of the predicted result. On ChIP-exo datasets, experimental results show that D-AEDNet has better performance than competing methods. Besides, we authenticate that Attention Gate can improve the interpretability of our model by ways of visualization analysis. Furthermore, we confirm that ability of D-AEDNet to learn TFs-DNA binding motifs outperform the state-of-the-art methods and availability of TF-MoDSW to discover biological sequence motifs in TFs-DNA interaction by conducting experiment on ChIP-seq datasets.

How to balance the bioinformatics data: pseudo-negative sampling.

BACKGROUND: Imbalanced datasets are commonly encountered in bioinformatics classification problems, that is, the number of negative samples is much larger than that of positive samples. Particularly, the data imbalance phenomena will make us underestimate the performance of the minority class of positive samples. Therefore, how to balance the bioinformatic data becomes a very challenging and difficult problem. RESULTS: In this study, we propose a new data sampling approach, called pseudo-negative sampling, which can be effectively applied to handle the case that: negative samples greatly dominate positive samples. Specifically, we design a supervised learning method based on a max-relevance min-redundancy criterion beyond Pearson correlation coefficient (MMPCC), which is used to choose pseudo-negative samples from the negative samples and view them as positive samples. In addition, MMPCC uses an incremental searching technique to select optimal pseudo-negative samples to reduce the computation cost. Consequently, the discovered pseudo-negative samples have strong relevance to positive samples and less redundancy to negative ones. CONCLUSIONS: To validate the performance of our method, we conduct experiments base on four UCI datasets and three real bioinformatics datasets. According to the experimental results, we clearly observe the performance of MMPCC is better than other sampling methods in terms of Sensitivity, Specificity, Accuracy and the Mathew's Correlation Coefficient. This reveals that the pseudo-negative samples are particularly helpful to solve the imbalance dataset problem. Moreover, the gain of Sensitivity from the minority samples with pseudo-negative samples grows with the improvement of prediction accuracy on all dataset.

Simultaneous denoising and super-resolution of optical coherence tomography images based on generative adversarial network.

Optical coherence tomography (OCT) has become a very promising diagnostic method in clinical practice, especially for ophthalmic diseases. However, speckle noise and low sampling rates have intensively reduced the quality of OCT images, which prevents the development of OCT-assisted diagnosis. Therefore, we propose a generative adversarial network-based approach (named SDSR-OCT) to simultaneously denoise and super-resolve OCT images. Moreover, we trained three different super-resolution models with different upscale factors (2× , 4× and 8×) to adapt to the corresponding downsampling rates. We also quantitatively and qualitatively compared our proposed method with some well-known algorithms. The experimental results show that our approach can effectively suppress speckle noise and can super-resolve OCT images at different scales.

3D conditional generative adversarial networks for high-quality PET image estimation at low dose.

Positron emission tomography (PET) is a widely used imaging modality, providing insight into both the biochemical and physiological processes of human body. Usually, a full dose radioactive tracer is required to obtain high-quality PET images for clinical needs. This inevitably raises concerns about potential health hazards. On the other hand, dose reduction may cause the increased noise in the reconstructed PET images, which impacts the image quality to a certain extent. In this paper, in order to reduce the radiation exposure while maintaining the high quality of PET images, we propose a novel method based on 3D conditional generative adversarial networks (3D c-GANs) to estimate the high-quality full-dose PET images from low-dose ones. Generative adversarial networks (GANs) include a generator network and a discriminator network which are trained simultaneously with the goal of one beating the other. Similar to GANs, in the proposed 3D c-GANs, we condition the model on an input low-dose PET image and generate a corresponding output full-dose PET image. Specifically, to render the same underlying information between the low-dose and full-dose PET images, a 3D U-net-like deep architecture which can combine hierarchical features by using skip connection is designed as the generator network to synthesize the full-dose image. In order to guarantee the synthesized PET image to be close to the real one, we take into account of the estimation error loss in addition to the discriminator feedback to train the generator network. Furthermore, a concatenated 3D c-GANs based progressive refinement scheme is also proposed to further improve the quality of estimated images. Validation was done on a real human brain dataset including both the normal subjects and the subjects diagnosed as mild cognitive impairment (MCI). Experimental results show that our proposed 3D c-GANs method outperforms the benchmark methods and achieves much better performance than the state-of-the-art methods in both qualitative and quantitative measures.

Fusing multi-scale information in convolution network for MR image super-resolution reconstruction.

BACKGROUND: Magnetic resonance (MR) images are usually limited by low spatial resolution, which leads to errors in post-processing procedures. Recently, learning-based super-resolution methods, such as sparse coding and super-resolution convolution neural network, have achieved promising reconstruction results in scene images. However, these methods remain insufficient for recovering detailed information from low-resolution MR images due to the limited size of training dataset. METHODS: To investigate the different edge responses using different convolution kernel sizes, this study employs a multi-scale fusion convolution network (MFCN) to perform super-resolution for MRI images. Unlike traditional convolution networks that simply stack several convolution layers, the proposed network is stacked by multi-scale fusion units (MFUs). Each MFU consists of a main path and some sub-paths and finally fuses all paths within the fusion layer. RESULTS: We discussed our experimental network parameters setting using simulated data to achieve trade-offs between the reconstruction performance and computational efficiency. We also conducted super-resolution reconstruction experiments using real datasets of MR brain images and demonstrated that the proposed MFCN has achieved a remarkable improvement in recovering detailed information from MR images and outperforms state-of-the-art methods. CONCLUSIONS: We have proposed a multi-scale fusion convolution network based on MFUs which extracts different scales features to restore the detail information. The structure of the MFU is helpful for extracting multi-scale information and making full-use of prior knowledge from a few training samples to enhance the spatial resolution.

Functional connectivity and activity of white matter in somatosensory pathways under tactile stimulations.

Functional MRI has proven to be effective in detecting neural activity in brain cortices on the basis of blood oxygenation level dependent (BOLD) contrast, but has relatively poor sensitivity for detecting neural activity in white matter. To demonstrate that BOLD signals in white matter are detectable and contain information on neural activity, we stimulated the somatosensory system and examined distributions of BOLD signals in related white matter pathways. The temporal correlation profiles and frequency contents of BOLD signals were compared between stimulation and resting conditions, and between relevant white matter fibers and background regions, as well as between left and right side stimulations. Quantitative analyses show that, overall, MR signals from white matter fiber bundles in the somatosensory system exhibited significantly greater temporal correlations with the primary sensory cortex and greater signal power during tactile stimulations than in a resting state, and were stronger than corresponding measurements for background white matter both during stimulations and in a resting state. The temporal correlation and signal power under stimulation were found to be twice those observed from the same bundle in a resting state, and bore clear relations with the side of stimuli. These indicate that BOLD signals in white matter fibers encode neural activity related to their functional roles connecting cortical volumes, which are detectable with appropriate methods.

Global denoising for 3D MRI.

BACKGROUND: Denoising is the primary preprocessing step for subsequent application of MRI. However, most commonly-used patch-based denoising methods are heavily dependent on the degree of patch matching. Due to the large number of voxels in the 3D MRI dataset, the procedure of searching sufficient similarity patches was limited by the empirical compromising between computational efficiency and estimation accuracy, and cannot fulfill the application in multimodal MRI dataset with different SNR and resolutions. METHODS: In this study, we propose a modified global filtering framework for 3D MRI. For each denoising voxel, the similarity weighting matrix is computed using the reference patch and other patches from the whole dataset. This large weighting matrix is then approximated using the k-means clustering Nyström method to achieve computational viability. RESULTS: Experiments on both synthetic and in vivo MRI datasets demonstrated that the proposed adaptive Nyström low-rank approximation could achieve competitive estimation compared with exact global filter while reducing the sampling rate by four orders of magnitude. In addition, the corresponding global filter improved patches-based method in both spatial and transform domain. CONCLUSION: We propose a global denoising framework for 3D MRI which extracts information from the entire dataset to restore each voxel. This large weighting matrix of the global filter is approximated using Nyström low-rank approximation with an adaptive k-means clustering sampling scheme, which significantly reduce the sampling rate as well as the running time. The proposed method is capable of denoising in multimodal MRI dataset and can be used to improve currently used patch-based methods.

Few-view image reconstruction with fractional-order total variation.

This work presents a novel computed tomography (CT) reconstruction method for the few-view problem based on fractional calculus. To overcome the disadvantages of the total variation minimization method, we propose a fractional-order total variation-based image reconstruction method in this paper. The presented model adopts fractional-order total variation instead of traditional total variation. Different from traditional total variation, fractional-order total variation is derived by considering more neighboring image voxels such that the corresponding weights can be adaptively determined by the model, thus suppressing the over-smoothing effect. The discretization scheme of the fractional-order model is also given. Numerical and clinical experiments demonstrate that our method achieves better performance than existing reconstruction methods, including filtered back projection (FBP), the total variation-based projections onto convex sets method (TV-POCS), and soft-threshold filtering (STH).

Accurate sparse-projection image reconstruction via nonlocal TV regularization.

Sparse-projection image reconstruction is a useful approach to lower the radiation dose; however, the incompleteness of projection data will cause degeneration of imaging quality. As a typical compressive sensing method, total variation has obtained great attention on this problem. Suffering from the theoretical imperfection, total variation will produce blocky effect on smooth regions and blur edges. To overcome this problem, in this paper, we introduce the nonlocal total variation into sparse-projection image reconstruction and formulate the minimization problem with new nonlocal total variation norm. The qualitative and quantitative analyses of numerical as well as clinical results demonstrate the validity of the proposed method. Comparing to other existing methods, our method more efficiently suppresses artifacts caused by low-rank reconstruction and reserves structure information better.

Prior Knowledge-guided Triple-Domain Transformer-GAN for Direct PET Reconstruction from Low-Count Sinograms.

To obtain high-quality positron emission tomography (PET) images while minimizing radiation exposure, numerous methods have been dedicated to acquiring standard-count PET (SPET) from low-count PET (LPET). However, current methods have failed to take full advantage of the different emphasized information from multiple domains, i.e., the sinogram, image, and frequency domains, resulting in the loss of crucial details. Meanwhile, they overlook the unique inner-structure of the sinograms, thereby failing to fully capture its structural characteristics and relationships. To alleviate these problems, in this paper, we proposed a prior knowledge-guided transformer-GAN that unites triple domains of sinogram, image, and frequency to directly reconstruct SPET images from LPET sinograms, namely PK-TriDo. Our PK-TriDo consists of a Sinogram Inner-Structure-based Denoising Transformer (SISD-Former) to denoise the input LPET sinogram, a Frequency-adapted Image Reconstruction Transformer (FaIR-Former) to reconstruct high-quality SPET images from the denoised sinograms guided by the image domain prior knowledge, and an Adversarial Network (AdvNet) to further enhance the reconstruction quality via adversarial training. Specifically tailored for the PET imaging mechanism, we injected a sinogram embedding module that partitions the sinograms by rows and columns to obtain 1D sequences of angles and distances to faithfully preserve the inner-structure of the sinograms. Moreover, to mitigate high-frequency distortions and enhance reconstruction details, we integrated global-local frequency parsers (GLFPs) into FaIR-Former to calibrate the distributions and proportions of different frequency bands, thus compelling the network to preserve high-frequency details. Evaluations on three datasets with different dose levels and imaging scenarios demonstrated that our PK-TriDo outperforms the state-of-the-art methods.

Promoting fast MR imaging pipeline by full-stack AI.

Magnetic resonance imaging (MRI) is a widely used imaging modality in clinics for medical disease diagnosis, staging, and follow-up. Deep learning has been extensively used to accelerate k-space data acquisition, enhance MR image reconstruction, and automate tissue segmentation. However, these three tasks are usually treated as independent tasks and optimized for evaluation by radiologists, thus ignoring the strong dependencies among them; this may be suboptimal for downstream intelligent processing. Here, we present a novel paradigm, full-stack learning (FSL), which can simultaneously solve these three tasks by considering the overall imaging process and leverage the strong dependence among them to further improve each task, significantly boosting the efficiency and efficacy of practical MRI workflows. Experimental results obtained on multiple open MR datasets validate the superiority of FSL over existing state-of-the-art methods on each task. FSL has great potential to optimize the practical workflow of MRI for medical diagnosis and radiotherapy.

3D multi-modality Transformer-GAN for high-quality PET reconstruction.

Positron emission tomography (PET) scans can reveal abnormal metabolic activities of cells and provide favorable information for clinical patient diagnosis. Generally, standard-dose PET (SPET) images contain more diagnostic information than low-dose PET (LPET) images but higher-dose scans can also bring higher potential radiation risks. To reduce the radiation risk while acquiring high-quality PET images, in this paper, we propose a 3D multi-modality edge-aware Transformer-GAN for high-quality SPET reconstruction using the corresponding LPET images and T1 acquisitions from magnetic resonance imaging (T1-MRI). Specifically, to fully excavate the metabolic distributions in LPET and anatomical structural information in T1-MRI, we first use two separate CNN-based encoders to extract local spatial features from the two modalities, respectively, and design a multimodal feature integration module to effectively integrate the two kinds of features given the diverse contributions of features at different locations. Then, as CNNs can describe local spatial information well but have difficulty in modeling long-range dependencies in images, we further apply a Transformer-based encoder to extract global semantic information in the input images and use a CNN decoder to transform the encoded features into SPET images. Finally, a patch-based discriminator is applied to ensure the similarity of patch-wise data distribution between the reconstructed and real images. Considering the importance of edge information in anatomical structures for clinical disease diagnosis, besides voxel-level estimation error and adversarial loss, we also introduce an edge-aware loss to retain more edge detail information in the reconstructed SPET images. Experiments on the phantom dataset and clinical dataset validate that our proposed method can effectively reconstruct high-quality SPET images and outperform current state-of-the-art methods in terms of qualitative and quantitative metrics.

SFT-Net: A Network for Detecting Fatigue From EEG Signals by Combining 4D Feature Flow and Attention Mechanism.

Fatigued driving is a leading cause of traffic accidents, and accurately predicting driver fatigue can significantly reduce their occurrence. However, modern fatigue detection models based on neural networks often face challenges such as poor interpretability and insufficient input feature dimensions. This paper proposes a novel Spatial-Frequency-Temporal Network (SFT-Net) method for detecting driver fatigue using electroencephalogram (EEG) data. Our approach integrates EEG signals' spatial, frequency, and temporal information to improve recognition performance. We transform the differential entropy of five frequency bands of EEG signals into a 4D feature tensor to preserve these three types of information. An attention module is then used to recalibrate the spatial and frequency information of each input 4D feature tensor time slice. The output of this module is fed into a depthwise separable convolution (DSC) module, which extracts spatial and frequency features after attention fusion. Finally, long short-term memory (LSTM) is used to extract the temporal dependence of the sequence, and the final features are output through a linear layer. We validate the effectiveness of our model on the SEED-VIG dataset, and experimental results demonstrate that SFT-Net outperforms other popular models for EEG fatigue detection. Interpretability analysis supports the claim that our model has a certain level of interpretability. Our work addresses the challenge of detecting driver fatigue from EEG data and highlights the importance of integrating spatial, frequency, and temporal information. Codes are available at https://github.com/wangkejie97/SFT-Net.

Unsupervised Domain Adaptive Dose Prediction via Cross-Attention Transformer and Target-Specific Knowledge Preservation.

Radiotherapy is one of the leading treatments for cancer. To accelerate the implementation of radiotherapy in clinic, various deep learning-based methods have been developed for automatic dose prediction. However, the effectiveness of these methods heavily relies on the availability of a substantial amount of data with labels, i.e. the dose distribution maps, which cost dosimetrists considerable time and effort to acquire. For cancers of low-incidence, such as cervical cancer, it is often a luxury to collect an adequate amount of labeled data to train a well-performing deep learning (DL) model. To mitigate this problem, in this paper, we resort to the unsupervised domain adaptation (UDA) strategy to achieve accurate dose prediction for cervical cancer (target domain) by leveraging the well-labeled high-incidence rectal cancer (source domain). Specifically, we introduce the cross-attention mechanism to learn the domain-invariant features and develop a cross-attention transformer-based encoder to align the two different cancer domains. Meanwhile, to preserve the target-specific knowledge, we employ multiple domain classifiers to enforce the network to extract more discriminative target features. In addition, we employ two independent convolutional neural network (CNN) decoders to compensate for the lack of spatial inductive bias in the pure transformer and generate accurate dose maps for both domains. Furthermore, to enhance the performance, two additional losses, i.e. a knowledge distillation loss (KDL) and a domain classification loss (DCL), are incorporated to transfer the domain-invariant features while preserving domain-specific information. Experimental results on a rectal cancer dataset and a cervical cancer dataset have demonstrated that our method achieves the best quantitative results with [Formula: see text], [Formula: see text], and HI of 1.446, 1.231, and 0.082, respectively, and outperforms other methods in terms of qualitative assessment.

A Transformer-Embedded Multi-Task Model for Dose Distribution Prediction.

Radiation therapy is a fundamental cancer treatment in the clinic. However, to satisfy the clinical requirements, radiologists have to iteratively adjust the radiotherapy plan based on experience, causing it extremely subjective and time-consuming to obtain a clinically acceptable plan. To this end, we introduce a transformer-embedded multi-task dose prediction (TransMTDP) network to automatically predict the dose distribution in radiotherapy. Specifically, to achieve more stable and accurate dose predictions, three highly correlated tasks are included in our TransMTDP network, i.e. a main dose prediction task to provide each pixel with a fine-grained dose value, an auxiliary isodose lines prediction task to produce coarse-grained dose ranges, and an auxiliary gradient prediction task to learn subtle gradient information such as radiation patterns and edges in the dose maps. The three correlated tasks are integrated through a shared encoder, following the multi-task learning strategy. To strengthen the connection of the output layers for different tasks, we further use two additional constraints, i.e. isodose consistency loss and gradient consistency loss, to reinforce the match between the dose distribution features generated by the auxiliary tasks and the main task. Additionally, considering many organs in the human body are symmetrical and the dose maps present abundant global features, we embed the transformer into our framework to capture the long-range dependencies of the dose maps. Evaluated on an in-house rectum cancer dataset and a public head and neck cancer dataset, our method gains superior performance compared with the state-of-the-art ones. Code is available at https://github.com/luuuwen/TransMTDP.

Facial Expression Recognition with Contrastive Learning and Uncertainty-Guided Relabeling.

Facial expression recognition (FER) plays a vital role in the field of human-computer interaction. To achieve automatic FER, various approaches based on deep learning (DL) have been presented. However, most of them lack for the extraction of discriminative expression semantic information and suffer from the problem of annotation ambiguity. In this paper, we propose an elaborately designed end-to-end recognition network with contrastive learning and uncertainty-guided relabeling, to recognize facial expressions efficiently and accurately, as well as to alleviate the impact of annotation ambiguity. Specifically, a supervised contrastive loss (SCL) is introduced to promote inter-class separability and intra-class compactness, thus helping the network extract fine-grained discriminative expression features. As for the annotation ambiguity problem, we present an uncertainty estimation-based relabeling module (UERM) to estimate the uncertainty of each sample and relabel the unreliable ones. In addition, to deal with the padding erosion problem, we embed an amending representation module (ARM) into the recognition network. Experimental results on three public benchmarks demonstrate that our proposed method facilitates the recognition performance remarkably with 90.91% on RAF-DB, 88.59% on FERPlus and 61.00% on AffectNet, outperforming current state-of-the-art (SOTA) FER methods. Code will be available at http//github.com/xiaohu-run/fer_supCon.

DeSeg: auto detector-based segmentation for brain metastases.

Delineation of brain metastases (BMs) is a paramount step in stereotactic radiosurgery treatment. Clinical practice has specific expectation on BM auto-delineation that the method is supposed to avoid missing of small lesions and yield accurate contours for large lesions. In this study, we propose a novel coarse-to-fine framework, named detector-based segmentation (DeSeg), to incorporate object-level detection into pixel-wise segmentation so as to meet the clinical demand. DeSeg consists of three components: a center-point-guided single-shot detector to localize the potential lesion regions, a multi-head U-Net segmentation model to refine contours, and a data cascade unit to connect both tasks smoothly. Performance on tiny lesions is measured by the object-based sensitivity and positive predictive value (PPV), while that on large lesions is quantified by dice similarity coefficient (DSC), average symmetric surface distance (ASSD) and 95% Hausdorff distance (HD95). Besides, computational complexity is also considered to study the potential of method in real-time processing. This study retrospectively collected 240 BM patients with Gadolinium injected contrast-enhanced T1-weighted magnetic resonance imaging (T1c-MRI), which were randomly split into training, validating and testing datasets (192, 24 and 24 scans, respectively). The lesions in the testing dataset were further divided into two groups based on the volume size (smallS: ≤1.5 cc,N= 88; largeL: > 1.5 cc,N= 15). On average, DeSeg yielded a sensitivity of 0.91 and a PPV of 0.77 on S group, and a DSC of 0.86, an ASSD 0f 0.76 mm and a HD95 of 2.31 mm onLgroup. The results indicated that DeSeg achieved leading sensitivity and PPV for tiny lesions as well as segmentation metrics for large ones. After our clinical validation, DeSeg showed competitive segmentation performance while kept faster processing speed comparing with existing 3D models.