Binary Cross Entropy Research Articles

SAST: A self-attention based method for skill translation in T-shaped expert finding

Nowadays, organizations are seeking professionals who can both excel in their areas of expertise and collaborate effectively across different disciplines. This demand has given rise to the concept of T-shaped experts who possess a deep understanding of one topic domain and a broad knowledge of several others. This combination allows these professionals to be more creative, flexible, and adaptable in problem-solving by leveraging their diverse perspectives and experiences. To find T-shaped experts in any skill area, we need to measure how deep and wide their knowledge is. In this paper, we present a novel translation-based method to estimate each user's depth of knowledge in a given skill area. The proposed method leverages a self-attention-based multi-label classification network to identify the most relevant translations for each skill that belongs to the given skill area. We utilize two new methods based on binary cross-entropy and focal loss to determine whether a user's expertise shape matches the T-shaped. We evaluate the proposed method using the standard benchmark datasets. The experimental results on three collections of the StackOverflow dataset demonstrate the superiority of the proposed methods in comparison with existing baselines.

ResACEUnet: An Improved Transformer Unet Model for 3D Seismic Fault Detection

AbstractDetecting fault constitutes a pivotal aspect of seismic interpretation, significantly influencing the outcomes of petroleum and gas exploration. As artificial intelligence advances, convolutional neural network (CNN) has proven effective in detecting faults in seismic interpretation. Nevertheless, the receptive field of a convolutional layer within CNN is inherently limited, focusing on extracting local features, which lead to the detection of fewer and discontinuous fault features. In this study, integrating the local feature extraction capabilities of CNN with the global feature extraction prowess of transformer, we proposed a U‐shaped hybrid architecture model named ResACEUnet (Attention‐Convolution Unet with Efficient block) to detect fault of three‐dimensional (3D) seismic data. In ResACEUnet, we introduced a module called ACE block, which integrates convolution and attention mechanisms. This module enabled the model to simultaneously extract local features and model global contextual information, capturing more accurate fault features. In addition, we utilized a joint loss function named BCEDice loss, which composed of BCE (binary cross‐entropy) loss and dice loss to tackle the challenge of imbalanced positive and negative samples. The model was trained on a synthetic data set, with a range of data augmentation techniques were employed to bolster its generalization capabilities and robustness. We implemented our proposed method on the offshore F3 seismic data from the Netherlands and seismic data from Kerry3D and Parihaka in New Zealand. Compared to conventional popular models such as Unet, ResUnet, and SwinUnetR, ResACEUnet demonstrated superior capabilities in capturing more features and identifying fault with higher accuracy and continuity.

Enhancing Brain Stroke Detection: A Novel Deep Neural Network with Weighted Binary Cross Entropy Training

Enhancing Brain Stroke Detection: A Novel Deep Neural Network with Weighted Binary Cross Entropy Training

Enhancing soccer pass receiver prediction in broadcast images through advanced deep learning techniques: A comprehensive study on model optimization and performance evaluation

In this study, we present a graph neural network (GNN) model specifically designed for football pass receiver prediction in Broadcast Images is presented in this study. Important node properties, including ball possession indicators, hot-encoded team values, and normalized ground placements, are incorporated into the model along with a careful weighting of edges to account for player distances. With weighted BCE loss used to overcome class imbalance, its architecture consists of a linear layer, numerous GNN Message Passing layers, a SoftMax activation, and binary cross-entropy (BCE) loss for training. Of 206 examples, 101 valid predictions were made, indicating a predictive accuracy of 0.50 according to the evaluation data. Comparative analyzes show that GAT-V2 (0.85) and GAT (0.63) perform better in terms of optimization and accuracy, respectively. The effectiveness in recognizing football pass receivers is demonstrated in this paper, highlighting developments in computer vision applications for sports analytics.

Enhancing Road Crack Localization for Sustainable Road Safety Using HCTNet

Road crack detection is crucial for maintaining and inspecting civil infrastructure, as cracks can pose a potential risk for sustainable road safety. Traditional methods for pavement crack detection are labour-intensive and time-consuming. In recent years, computer vision approaches have shown encouraging results in automating crack localization. However, the classical convolutional neural network (CNN)-based approach lacks global attention to the spatial features. To improve the crack localization in the road, we designed a vision transformer (ViT) and convolutional neural networks (CNNs)-based encoder and decoder. In addition, a gated-attention module in the decoder is designed to focus on the upsampling process. Furthermore, we proposed a hybrid loss function using binary cross-entropy and Dice loss to evaluate the model’s effectiveness. Our method achieved a recall, F1-score, and IoU of 98.54%, 98.07%, and 98.72% and 98.27%, 98.69%, and 98.76% on the Crack500 and Crack datasets, respectively. Meanwhile, on the proposed dataset, these figures were 96.89%, 97.20%, and 97.36%.

Transfer-Learning-Enhanced Regression Generative Adversarial Networks for Optimal eVTOL Takeoff Trajectory Prediction

Electric vertical takeoff and landing (eVTOL) aircraft represent a crucial aviation technology to transform future transportation systems. The unique characteristics of eVTOL aircraft include reduced noise, low pollutant emission, efficient operating cost, and flexible maneuverability, which in the meantime pose critical challenges to advanced power retention techniques. Thus, optimal takeoff trajectory design is essential due to immense power demands during eVTOL takeoffs. Conventional design optimizations, however, adopt high-fidelity simulation models in an iterative manner resulting in a computationally intensive mechanism. In this work, we implement a surrogate-enabled inverse mapping optimization architecture, i.e., directly predicting optimal designs from design requirements (including flight conditions and design constraints). A trained inverse mapping surrogate performs real-time optimal eVTOL takeoff trajectory predictions with no need for running optimizations; however, one training sample requires one design optimization in this inverse mapping setup. The excessive training cost of inverse mapping and the characteristics of optimal eVTOL takeoff trajectories necessitate the development of the regression generative adversarial network (regGAN) surrogate. We propose to further enhance regGAN predictive performance through the transfer learning (TL) technique, creating a scheme termed regGAN-TL. In particular, the proposed regGAN-TL scheme leverages the generative adversarial network (GAN) architecture consisting of a generator network and a discriminator network, with a combined loss of the mean squared error (MSE) and binary cross-entropy (BC) losses, for regression tasks. In this work, the generator employs design requirements as input and produces optimal takeoff trajectory profiles, while the discriminator differentiates the generated profiles and real optimal profiles in the training set. The combined loss facilitates the generator training in the dual aspects: the MSE loss targets minimum differences between generated profiles and training counterparts, while the BC loss drives the generated profiles to share analogous patterns with the training set. We demonstrated the utility of regGAN-TL on optimal takeoff trajectory designs for the Airbus A3 Vahana and compared its performance against representative surrogates, including the multi-output Gaussian process, the conditional GAN, and the vanilla regGAN. Results showed that regGAN-TL reached the 99.5% generalization accuracy threshold with only 200 training samples while the best reference surrogate required 400 samples. The 50% reduction in training expense and reduced standard deviations of generalization accuracy achieved by regGAN-TL confirmed its outstanding predictive performance and broad engineering application potential.

Weakly supervised instance segmentation via class double-activation maps and boundary localization

Weakly supervised instance segmentation based on image-level class labels has recently gained much attention, in which the primary key step is to generate the pseudo labels based on class activation maps (CAMs). Most methods adopt binary cross-entropy (BCE) loss to train the classification model. However, since BCE loss is not class mutually exclusive, activations among classes occur independently. Thus, not only do foreground classes are wrongly activated as background, but also incorrect activations among confusing classes are occurred in the foreground. To solve this problem, we propose the Class Double-Activation Map, called Double-CAM. Firstly, the vanilla CAM is extracted from the multi-label classifier and then fused with the output feature map of backbone. The enhanced feature map of each class is fed into the single-label classification branch with softmax cross-entropy (SCE) loss and entropy minimization module, from which the more accurate Double-CAM is extracted. It refines the vanilla CAM to improve the quality of pseudo labels. Secondly, to mine object edge cues from Double-CAM, we propose the Boundary Localization (BL) module to synthesize boundary annotations, so as to provide constraints for label propagation more explicitly without adding additional supervision. The quality of pseudo masks is also improved substantially with the addition of BL module. Finally, the generated pseudo labels are used to train fully supervised instance segmentation networks. The evaluations on VOC and COCO datasets show that our method achieves excellent performance, outperforming mainstream weakly supervised segmentation methods at the same supervisory level, even those that depend on stronger supervision.

RoUIE: A Method for Constructing Knowledge Graph of Power Equipment Based on Improved Universal Information Extraction

The current state evaluation of power equipment often focuses solely on changes in electrical quantities while neglecting basic equipment information as well as textual information such as system alerts, operation records, and defect records. Constructing a device-centric knowledge graph by extracting information from multiple sources related to power equipment is a valuable approach to enhance the intelligence level of asset management. Through the collection of pertinent authentic datasets, we have established a dataset for the state evaluation of power equipment, encompassing 35 types of relationships. To better suit the characteristics of concentrated relationship representations and varying lengths in textual descriptions, we propose a generative model called RoUIE, which is a method for constructing a knowledge graph of power equipment based on improved Universal Information Extraction (UIE). This model first utilizes a pre-trained language model based on rotational position encoding as the text encoder in the fine-tuning stage. Subsequently, we innovatively leverage the Distribution Focal Loss (DFL) to replace Binary Cross-Entropy Loss (BCE) as the loss function, further enhancing the model’s extraction performance. The experimental results demonstrate that compared to the UIE model and mainstream joint extraction benchmark models, RoUIE exhibits superior performance on the dataset we constructed. On a general Chinese dataset, the proposed model also outperforms baseline models, showcasing the model’s universal applicability.

Two-headed UNetEfficientNets for parallel execution of segmentation and classification of brain tumors: incorporating postprocessing techniques with connected component labelling.

The purpose of this study is to develop accurate and automated detection and segmentation methods for brain tumors, given their significant fatality rates, with aggressive malignant tumors like Glioblastoma Multiforme (GBM) having a five-year survival rate as low as 5 to 10%. This underscores the urgent need to improve diagnosis and treatment outcomes through innovative approaches in medical imaging and deep learning techniques. In this work, we propose a novel approach utilizing the two-headed UNetEfficientNets model for simultaneous segmentation and classification of brain tumors from Magnetic Resonance Imaging (MRI) images. The model combines the strengths of EfficientNets and a modified two-headed Unet model. We utilized a publicly available dataset consisting of 3064 brain MR images classified into three tumor classes: Meningioma, Glioma, and Pituitary. To enhance the training process, we performed 12 types of data augmentation on the training dataset. We evaluated the methodology using six deep learning models, ranging from UNetEfficientNet-B0 to UNetEfficientNet-B5, optimizing the segmentation and classification heads using binary cross entropy (BCE) loss with Dice and BCE with focal loss, respectively. Post-processing techniques such as connected component labeling (CCL) and ensemble models were applied to improve segmentation outcomes. The proposed UNetEfficientNet-B4 model achieved outstanding results, with an accuracy of 99.4% after postprocessing. Additionally, it obtained high scores for DICE (94.03%), precision (98.67%), and recall (99.00%) after post-processing. The ensemble technique further improved segmentation performance, with a global DICE score of 95.70% and Jaccard index of 91.20%. Our study demonstrates the high efficiency and accuracy of the proposed UNetEfficientNet-B4 model in the automatic and parallel detection and segmentation of brain tumors from MRI images. This approach holds promise for improving diagnosis and treatment planning for patients with brain tumors, potentially leading to better outcomes and prognosis.

Image Inpainting based on Dilated Neighborhood Attention

In response to the phenomenon that existing image restoration algorithms have blurry edges, incoherent textures, and lack clarity and delicacy for large-area missing images, a two-stage generative adversarial image restoration algorithm based on dilated neighborhood attention is proposed. This algorithm decouples image restoration into edge structure restoration and texture structure restoration, introduces a dilated neighborhood attention mechanism, and enhances the generator's focus on important information and structures in the image by constructing a residual attention network, thereby improving the perception and utilization of texture details, resulting in more realistic image views and finer texture details. This paper introduces the Binary Cross-Entropy with Logits loss function in the discriminators of the two stages, which can help the discriminator learn how to more effectively identify real and generated images, thus improving the overall network performance. The Ranger21 optimizer is introduced to accelerate learning without affecting generalization, addressing the problem of traditional optimizers systematically staying in poor initial states. The datasets used in this paper are Paris Street View and CelebA-HQ. Comparative experiments with other image restoration methods show that both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) have improved, and the larger the mask area, the more significant the improvement. Experiments prove that the images restored by the proposed algorithm have more reasonable structures and richer details, and the image restoration effect is superior.

A comparison framework for deep learning RFI detection algorithms

ABSTRACT In the realm of radio astronomy, the detection of radio frequency interference (RFI) stands as a pivotal pursuit. This study performs a novel comparative analysis of deep learning methodologies and introduces a novel transfer learning method, called fine-tuning. We compare various aspects and factors relating to this problem, including supervised fully convolutional network (FCN) architectures used within the literature, loss functions, regularization techniques, and training methodologies, to establish the most effective strategies for RFI detection. Moreover, the relationship between parameters, FLOPS, and inference times are examined. Fine-tuning involves pre-training models with low-quality aoflagger reference outputs, a very popular and accessible RFI flagging software package, and thereafter re-training the models with high-quality reference outputs. We utilize two data sets: real observations from LOFAR and simulated data from HERA. The mean squared error (MSE) loss function emerges as a robust performer if a high recall is desired. In contrast, the binary cross-entropy (BCE) loss function excels in generalization but falls short in classification performance. The Dice loss function emerges as the top performer, maximizing the F1 score and thereby serving as the choice for our further investigations. Notably, we highlight the important role of data quality and model capacity. In particular, we find that low-capacity models exhibit resilience when trained with low-quality flags from aoflagger, showcasing their ability to mitigate overfitting and overflagging tendencies. In contrast, high-capacity models excel when trained with high-quality flags. Fine-tuning proved to be an effective method to unlearn the overflagging tendencies of aoflagger, whilst requiring very little data.

On the Effectiveness of Sampled Softmax Loss for Item Recommendation

The learning objective plays a fundamental role to build a recommender system. Most methods routinely adopt either pointwise (e.g., binary cross-entropy) or pairwise (e.g., BPR) loss to train the model parameters, while rarely pay attention to softmax loss, which assumes the probabilities of all classes sum up to 1, due to its computational complexity when scaling up to large datasets or intractability for streaming data where the complete item space is not always available. The sampled softmax (SSM) loss emerges as an efficient substitute for softmax loss. Its special case, InfoNCE loss, has been widely used in self-supervised learning and exhibited remarkable performance for contrastive learning. Nonetheless, limited recommendation work uses the SSM loss as the learning objective. Worse still, none of them explores its properties thoroughly and answers “Does SSM loss suit for item recommendation?” and “What are the conceptual advantages of SSM loss, as compared with the prevalent losses?”, to the best of our knowledge. In this work, we aim at offering a better understanding of SSM for item recommendation. Specifically, we first theoretically reveal three model-agnostic advantages: (1) mitigating popularity bias, which is beneficial to long-tail recommendation; (2) mining hard negative samples, which offers informative gradients to optimize model parameters; and (3) maximizing the ranking metric, which facilitates top- K performance. However, based on our empirical studies, we recognize that the default choice of cosine similarity function in SSM limits its ability in learning the magnitudes of representation vectors. As such, the combinations of SSM with the models that also fall short in adjusting magnitudes (e.g., matrix factorization) may result in poor representations. One step further, we provide mathematical proof that message passing schemes in graph convolution networks can adjust representation magnitude according to node degree, which naturally compensates for the shortcoming of SSM. Extensive experiments on four benchmark datasets justify our analyses, demonstrating the superiority of SSM for item recommendation. Our implementations are available in both TensorFlow 1 and PyTorch. 2

A roundtrip probability estimation method for mechanical equipment fault detection under imbalanced samples

Aiming at high misdetection of mechanical faults under imbalanced samples, a roundtrip probability-based method is proposed. By roundtrip mapping between latent variables and real fault data, biased estimation of the probability distribution of real fault data is obtained. Further, virtual fault data are sampled according to such distribution to increase sample amount. For recognition of real and virtual data, loss function based on binary cross-entropy is designed. For reconstruction between fault data and its roundtrip mapped results, objective function based on mean square error is designed. Thus, it preserves boundary data and avoids too many virtual data in central area. Meanwhile, a strategy for eliminating abnormal samples is designed to reduce boundary deviation. For supporting the advantage of roundtrip, in-depth reasons for misdetection are analyzed from empirical risk and structural risk. Experiments on 30 benchmark imbalanced test sets show that fault detection rate increases after amount enhancement. Additionally, it is verified on blade cracking and bearing fault detection. Results show that F1 score increases from 0.485 to 0.51 and 0.725 to 0.775 for such two cases.

Image Segmentation Based On U-Net and Adjusted U-Nets

Image segmentation, the process of dividing an image into various areas is quite crucial recently. Numerous industries, including robotics, remote sensing, and medical imaging, have applications for this task. In recent years, deep learning techniques, especially U-shaped networks (U-nets), have shown remarkable success in solving image segmentation problems. This paper provides an overview of image segmentation using neural networks, introduces different types of adjusted U-nets used for this task, including the implementation of attention gates and the use of residual neural network as the encoder path based on the original encoder-decoder structure, and then uses U-net and adjusted U-nets to conduct image segmentation on the black sea sprat. The study uses dice similarity coefficient and binary cross entropy loss function to compare the model training results and further judges the functionality of the models by the predicted segmented images. According to the test results, the Res34-UNet with attention gates performs most efficiently in segmenting this image dataset, although it's more unstable compared to the basic U-Net.

Superpixel-wise contrast exploration for salient object detection

Salient object detection (SOD) methods typically consider SOD as a pixel-wise binary classification problem and utilize the binary cross-entropy (BCE) loss for optimization. However, the BCE loss ignores the global dependencies between pixels of the entire image, which is important for ensuring the accuracy and integrity of objects. To address this limitation explicitly, contrastive learning is introduced to enhance both the intra-pixel compactness of the foreground and inter-pixel separability between the foreground and background. Unfortunately, the random pixel sampling strategies of existing contrast algorithms cannot ensure that the sampled pixels cover all semantic contents across the entire image because of the lack of fine-grained semantic categories for pixels in SOD. Hence, in this study, a SuperPixel-wise Contrastive (SPCont) algorithm is proposed, which samples all superpixel centers to improve the diversity and representativeness of the sampled pixels. In addition, a SOD-specific hard pixel sampling strategy is designed to refine the edge details of the objects. An intra-image multi-level negative sampling approach is presented to increase the number of negative samples. Extensive experiments are conducted to demonstrate that incorporating the SPCont algorithm into the state-of-the-art network results in significant performance improvements. We expect that our work will open new paths for the application of contrastive learning to SOD.

Semantic segmentation of thermal defects in belt conveyor idlers using thermal image augmentation and U-Net-based convolutional neural networks

The belt conveyor (BC) is the main means of horizontal transportation of bulk materials at mining sites. The sudden fault in BC modules may cause unexpected stops in production lines. With the increasing number of applications of inspection mobile robots in condition monitoring (CM) of industrial infrastructure in hazardous environments, in this article we introduce an image processing pipeline for automatic segmentation of thermal defects in thermal images captured from BC idlers using a mobile robot. This study follows the fact that CM of idler temperature is an important task for preventing sudden breakdowns in BC system networks. We compared the performance of three different types of U-Net-based convolutional neural network architectures for the identification of thermal anomalies using a small number of hand-labeled thermal images. Experiments on the test data set showed that the attention residual U-Net with binary cross entropy as the loss function handled the semantic segmentation problem better than our previous research and other studied U-Net variations.

A joint autoencoder and classifier deep neural network for AD and MCI classification

AbstractIn this article, we present a new approach to distinguish progressive mild cognitively impaired (pMCI) subjects, who eventually develop Alzheimer's disease (AD) from stable MCI (sMCI) subjects whose situation does not deteriorate into AD. The proposed approach combines the discriminating capabilities of classifiers and representation learning capacities of autoencoders into a unified architecture, and is hence termed as joint autoencoder and classifier deep neural network (JACDNN). JACDNN employs a single classifier and multiple autoencoders that are trained together to perform pattern classification. The classifier in JACDNN is trained using standard approaches to distinguish between subject from different classes using the binary cross entropy loss. The autoencoders in JACDNN, regularizes individual layers in the network used for classification to learn representations useful for reconstructing a given input. The performance of JACDNN has been evaluated on several machine learning problems pertaining to dementia, namely AD versus cognitively normal (CN) subjects, AD versus sMCI, CN versus pMCI, and pMCI versus sMCI. These problems are targeted using two datasets. The first dataset consist of gray matter (GM) features of subjects and the second dataset consist of combination of GM and white matter (WM) features. It is observed that better classification results are obtained when the classifier is built on GM and WM as compared with GM features alone. Performance comparison of JACDNN with other existing approaches has been conducted for these problems. The results clearly indicate that JACDNN performs better than other existing approaches for these problems.

MULTI-LABEL CLASSIFICATION OF THORACIC DISEASES USING STRUCTURED DEEP LEARNING FRAMEWORK

The current medical scenario indicates that thoracic diseases are the primary cause of illnesses of human beings worldwide. COVID-19 outbreak results in the generation of tremendous Chest X-Ray (CXR) and Computed Tomography (CT) image data archives because CXR and CT imaging are the foremost effective screening tools that characterize patterns and features of pathologies present in the lungs. After careful observation, these chest radiographs show that images are often labeled with more than one pathology, that extends disease diagnosis problem to tedious multi-label classification task. So, this paper proposes a modified mish activation function and compound loss-based Separable Convolution Neural Network (SCNN) model that accomplishes image-level detection of multiple pathologies and COVID-19 from multi-modality chest radiographs. We investigate the power of separable convolution in the CNN network that interprets spatial and depth-wise dimensions of each pixel in CXR and CT images. A modified mish called as Swmish, a non-monotonic activation function, is introduced that preserves negative gradient flow of input and capture fine-grained detail of multiple regions. Moreover, a compound loss function of dice coefficient and binary cross entropy is applied that strongly optimizes the model during training. The generality of the proposed SCNN model is confirmed after conducting comprehensive experiments on multi-label CXR and COVID19-CT datasets. The proposed model gains an AUC score of 0.91 and loss of 0.13 on the multi-label CXR dataset and achieves a testing accuracy of 0.97, AUC score of 0.99 and a loss of 0.11 on the COVID19-CT dataset. These results show that the proposed model achieves highest evaluation metrics in classifying binary and multi-label thoracic abnormalities from multi-modality chest radiographs.

ASD-Net: a novel U-Net based asymmetric spatial-channel convolution network for precise kidney and kidney tumor image segmentation

Early intervention in tumors can greatly improve human survival rates. With the development of deep learning technology, automatic image segmentation has taken a prominent role in the field of medical image analysis. Manually segmenting kidneys on CT images is a tedious task, and due to the diversity of these images and varying technical skills of professionals, segmentation results can be inconsistent. To address this problem, a novel ASD-Net network is proposed in this paper for kidney and kidney tumor segmentation tasks. First, the proposed network employs newly designed Adaptive Spatial-channel Convolution Optimization (ASCO) blocks to capture anisotropic information in the images. Then, other newly designed blocks, i.e., Dense Dilated Enhancement Convolution (DDEC) blocks, are utilized to enhance feature propagation and reuse it across the network, thereby improving its segmentation accuracy. To allow the network to segment complex and small kidney tumors more effectively, the Atrous Spatial Pyramid Pooling (ASPP) module is incorporated in its middle layer. With its generalized pyramid feature, this module enables the network to better capture and understand context information at various scales within the images. In addition to this, the concurrent spatial and channel squeeze & excitation (scSE) attention mechanism is adopted to better comprehend and manage context information in the images. Additional encoding layers are also added to the base (U-Net) and connected to the original encoding layer through skip connections. The resultant enhanced U-Net structure allows for better extraction and merging of high-level and low-level features, further boosting the network’s ability to restore segmentation details. In addition, the combined Binary Cross Entropy (BCE)-Dice loss is utilized as the network's loss function. Experiments, conducted on the KiTS19 dataset, demonstrate that the proposed ASD-Net network outperforms the existing segmentation networks according to all evaluation metrics used, except for recall in the case of kidney tumor segmentation, where it takes the second place after Attention-UNet.Graphical

CellT-Net: A Composite Transformer Method for 2-D Cell Instance Segmentation.

Cell instance segmentation (CIS) via light microscopy and artificial intelligence (AI) is essential to cell and gene therapy-based health care management, which offers the hope of revolutionary health care. An effective CIS method can help clinicians to diagnose neurological disorders and quantify how well these deadly disorders respond to treatment. To address the CIS task challenged by dataset characteristics such as irregular morphology, variation in sizes, cell adhesion, and obscure contours, we propose a novel deep learning model named CellT-Net to actualize effective cell instance segmentation. In particular, the Swin transformer (Swin-T) is used as the basic model to construct the CellT-Net backbone, as the self-attention mechanism can adaptively focus on useful image regions while suppressing irrelevant background information. Moreover, CellT-Net incorporating Swin-T constructs a hierarchical representation and generates multi-scale feature maps that are suitable for detecting and segmenting cells at different scales. A novel composite style named cross-level composition (CLC) is proposed to build composite connections between identical Swin-T models in the CellT-Net backbone and generate more representational features. The earth mover's distance (EMD) loss and binary cross entropy loss are used to train CellT-Net and actualize the precise segmentation of overlapped cells. The LiveCELL and Sartorius datasets are utilized to validate the model effectiveness, and the results demonstrate that CellT-Net can achieve better model performance for dealing with the challenges arising from the characteristics of cell datasets than state-of-the-art models.