Cross-domain Transfer Learning Research Articles

In-Domain Transfer Learning Strategy for Tumor Detection on Brain MRI.

Transfer learning has gained importance in areas where there is a labeled data shortage. However, it is still controversial as to what extent natural image datasets as pre-training sources contribute scientifically to success in different fields, such as medical imaging. In this study, the effect of transfer learning for medical object detection was quantitatively compared using natural and medical image datasets. Within the scope of this study, transfer learning strategies based on five different weight initialization methods were discussed. A natural image dataset MS COCO and brain tumor dataset BraTS 2020 were used as the transfer learning source, and Gazi Brains 2020 was used for the target. Mask R-CNN was adopted as a deep learning architecture for its capability to effectively handle both object detection and segmentation tasks. The experimental results show that transfer learning from the medical image dataset was found to be 10% more successful and showed 24% better convergence performance than the MS COCO pre-trained model, although it contains fewer data. While the effect of data augmentation on the natural image pre-trained model was 5%, the same domain pre-trained model was measured as 2%. According to the most widely used object detection metric, transfer learning strategies using MS COCO weights and random weights showed the same object detection performance as data augmentation. The performance of the most effective strategies identified in the Mask R-CNN model was also tested with YOLOv8. Results showed that even if the amount of data is less than the natural dataset, in-domain transfer learning is more efficient than cross-domain transfer learning. Moreover, this study demonstrates the first use of the Gazi Brains 2020 dataset, which was generated to address the lack of labeled and qualified brain MRI data in the medical field for in-domain transfer learning. Thus, knowledge transfer was carried out from the deep neural network, which was trained with brain tumor data and tested on a different brain tumor dataset.

Cross-domain knowledge transfer based parallel-cascaded multi-scale attention network for limited view reconstruction in projection magnetic particle imaging

Projection magnetic particle imaging (MPI) can significantly improve the temporal resolution of three-dimensional (3D) imaging compared to that using traditional point by point scanning. However, the dense view of projections required for tomographic reconstruction limits the scope of temporal resolution optimization. The solution to this problem in computed tomography (CT) is using limited view projections (sparse view or limited angle) for reconstruction, which can be divided into: completing the limited view sinogram and image post-processing for streaking artifacts caused by insufficient projections. Benefiting from large-scale CT datasets, both categories of deep learning-based methods have achieved tremendous progress; yet, there is a data scarcity limitation in MPI. We propose a cross-domain knowledge transfer learning strategy that can transfer the prior knowledge of the limited view learned by the model in CT to MPI, which can help reduce the network requirements for real MPI data. In addition, the size of the imaging target affects the scale of the streaking artifacts caused by insufficient projections. Therefore, we propose a parallel-cascaded multi-scale attention module that allows the network to adaptively identify streaking artifacts at different scales. The proposed method was evaluated on real phantom and in vivo mouse data, and it significantly outperformed several advanced limited view methods. The streaking artifacts caused by an insufficient number of projections can be overcome using the proposed method.

Skin lesion segmentation using two-phase cross-domain transfer learning framework

Background and ObjectiveDeep learning (DL) models have been used for medical imaging for a long time but they did not achieve their full potential in the past because of insufficient computing power and scarcity of training data. In recent years, we have seen substantial growth in DL networks because of improved technology and an abundance of data. However, previous studies indicate that even a well-trained DL algorithm may struggle to generalize data from multiple sources because of domain shifts. Additionally, ineffectiveness of basic data fusion methods, complexity of segmentation target and low interpretability of current DL models limit their use in clinical decisions. To meet these challenges, we present a new two-phase cross-domain transfer learning system for effective skin lesion segmentation from dermoscopic images. MethodsOur system is based on two significant technical inventions. We examine a two- phase cross-domain transfer learning approach, including model-level and data-level transfer learning, by fine-tuning the system on two datasets, MoleMap and ImageNet. We then present nSknRSUNet, a high-performing DL network, for skin lesion segmentation using broad receptive fields and spatial edge attention feature fusion. We examine the trained model's generalization capabilities on skin lesion segmentation to quantify these two inventions. We cross-examine the model using two skin lesion image datasets, MoleMap and HAM10000, obtained from varied clinical contexts. ResultsAt data-level transfer learning for the HAM10000 dataset, the proposed model obtained 94.63% of DSC and 99.12% accuracy. In cross-examination at data-level transfer learning for the Molemap dataset, the proposed model obtained 93.63% of DSC and 97.01% of accuracy. ConclusionNumerous experiments reveal that our system produces excellent performance and improves upon state-of-the-art methods on both qualitative and quantitative measures.

Rainstorm-Induced Emergency Recognition from Citizens’ Communications Based on Spatial Feature Extraction and Transfer Learning

The rapid recognition of rainstorm-induced emergencies (e.g., building and road inundation, facility damage, and trapped citizens) is vital to timely disaster response. One big challenge that limits the performance of emergency recognition is the data imbalance between different emergency domains. The present study aims to develop an effective cross-domain transfer learning framework for rainstorm-induced emergency recognition based on the text reports provided by citizens. The critical component of the framework is the use of joint distribution adaption (JDA) analysis embedded in a discriminative feature mapping procedure, which transfers rich knowledge learned from large-scale datasets to the learning task from small emergence data. Considering the feature incompleteness that is caused by short text length, a basic probability assignment function is constructed and applied to extract important spatial features for rainstorm emergency recognition, with an improved marginal Fisher analysis being adopted to optimize cross-domain text feature representation. The proposed scheme is validated using the empirical data of ten emergency classes from Wuhan City, China. Our experimental results show that the proposed method could significantly address data imbalance and thus help achieve high recognition performance through domain knowledge complementation. Meanwhile, the use of various spatial features is proved to be effective in tackling missing features. This scheme can be further developed into smart systems for rainstorm disaster response with reasonable performance and imbalanced sample sizes.

Identifying Solitary Granulomatous Nodules from Solid Lung Adenocarcinoma: Exploring Robust Image Features with Cross-Domain Transfer Learning.

This study aimed to find suitable source domain data in cross-domain transfer learning to extract robust image features. Then, a model was built to preoperatively distinguish lung granulomatous nodules (LGNs) from lung adenocarcinoma (LAC) in solitary pulmonary solid nodules (SPSNs). Data from 841 patients with SPSNs from five centres were collected retrospectively. First, adaptive cross-domain transfer learning was used to construct transfer learning signatures (TLS) under different source domain data and conduct a comparative analysis. The Wasserstein distance was used to assess the similarity between the source domain and target domain data in cross-domain transfer learning. Second, a cross-domain transfer learning radiomics model (TLRM) combining the best performing TLS, clinical factors and subjective CT findings was constructed. Finally, the performance of the model was validated through multicentre validation cohorts. Relative to other source domain data, TLS based on lung whole slide images as source domain data (TLS-LW) had the best performance in all validation cohorts (AUC range: 0.8228-0.8984). Meanwhile, the Wasserstein distance of TLS-LW was 1.7108, which was minimal. Finally, TLS-LW, age, spiculated sign and lobulated shape were used to build the TLRM. In all validation cohorts, The AUC ranges were 0.9074-0.9442. Compared with other models, decision curve analysis and integrated discrimination improvement showed that TLRM had better performance. The TLRM could assist physicians in preoperatively differentiating LGN from LAC in SPSNs. Furthermore, compared with other images, cross-domain transfer learning can extract robust image features when using lung whole slide images as source domain data and has a better effect.

Annotation Cost Minimization for Ultrasound Image Segmentation using Cross-domain Transfer Learning.

Deep learning techniques can help minimize inter-physician analysis variability and the medical expert workloads, thereby enabling more accurate diagnoses. However, their implementation requires large-scale annotated dataset whose acquisition incurs heavy time and human-expertise costs. Hence, to significantly minimize the annotation cost, this study presents a novel framework that enables the deployment of deep learning methods in ultrasound (US) image segmentation requiring only very limited manually annotated samples. We propose SegMix, a fast and efficient approach that exploits a segment-paste-blend concept to generate large number of annotated samples based on a few manually acquired labels. Besides, a series of US-specific augmentation strategies built upon image enhancement algorithms are introduced to make maximum use of the available limited number of manually delineated images. The feasibility of the proposed framework is validated on the left ventricle (LV) segmentation and fetal head (FH) segmentation tasks, respectively. Experimental results demonstrate that using only 10 manually annotated images, the proposed framework can achieve a Dice and JI of 82.61% and 83.92%, and 88.42% and 89.27% for LV segmentation and FH segmentation, respectively. Compared with training using the entire training set, there is over 98% of annotation cost reduction while achieving comparable segmentation performance. This indicates that the proposed framework enables satisfactory deep leaning performance when very limited number of annotated samples is available. Therefore, we believe that it can be a reliable solution for annotation cost reduction in medical image analysis.

Chest X-Ray Quality Assessment Method With Medical Domain Knowledge Fusion

Digital X-ray radiography is widely used in clinical diagnosis. High quality chest X-ray is conducive to the accurate diagnosis of diseases by clinicians. However, the quality assessment of the chest X-ray images mainly depends on the subjective evaluation of doctors, and the results are influenced by the skill level and experience of the evaluators, involving many issues such as heavy workload and various uncertain factors in such subjective judgment. In this paper, we propose a chest X-ray quality assessment method that combines image-text contrastive learning with medical domain knowledge fusion. Based on pretraining the model from contrastive text-image pairs, large-scale real clinical chest X-ray and diagnostic report text information are fused, and the model is fine-tuned to achieve cross domain transfer learning. While improving the prediction accuracy of the algorithm, the cost of massive sample data annotation is avoided. The local visual patch features of the X-ray images are aligned with multiple text features to ensure that the visual features contain more fine-grained image information. Theoretical analysis and experimental results show that the contrastive learning algorithm based on the fusion of triplet information in medical knowledge graph and chest X-ray multi-modal data has achieved good performance in terms of accuracy. In addition, the method proposed in this paper can be easily extended to complete other tasks such as medical image multi-lesion segmentation and disease progression prediction.

TomFusioNet: A tomato crop analysis framework for mobile applications using the multi-objective optimization based late fusion of deep models and background elimination

Crop disease is a critical concern for the farmers which needs to be addressed to mitigate production loss. Multifarious frameworks have been proposed for the rapid recognition of tomato crop diseases, however, their practical deployment is a daunting task. The existing models suffer from performance issues such as overfitting and gradient vanishing. To alleviate these issues, this research proposes an end-to-end tomato crop analysis framework, TomFusioNet, for mobile applications. For feature extraction, the late-fusion technique is leveraged by aggregating the results of modified cross-domain transfer learning models using the maximum likelihood prediction strategy. Multi-layer Perceptron models are utilized as meta-learners. TomFusioNet’s pipeline comprises two modules, namely, DeepRec and DeepPred. DeepRec provides preliminary disease recognition results while DeepPred further identifies the type of illness in crop. Logic gate mapping is employed to reduce unnecessary wastage of mobile computation. The hyperparameters of DeepPred are tuned using the multi-objective optimization based non-dominated sorting genetic algorithm II for performance enhancement. We highlighted the significance of feature relevancy; therefore, a hue, saturation and value (HSV) color model-based background elimination technique is also proposed. TomFusioNet can be incorporated in a smartphone app, conceptualized for remote crop monitoring. The proposed DeepRec and DeepPred models achieved an average accuracy of 99.93% and 98.32%, respectively. According to experimental results, TomFusioNet, which is built on NSGA-II, beats state-of-the-art models with a convergence loss of 0.021 and an AUC value of 99.10%. The proposed app framework’s functioning does not require any human in loop. Furthermore, the latency of framework in real-time is less than 2 s, hence, it is effective for rapid tomato crop analysis.

Towards rainstorm event identification: A transfer learning framework using citizen-report texts and multi-source spatial data

Urban rainstorms are often accompanied by various hazardous events (e.g., water inundation and small-scale debris flow) and subsequent consequential events (e.g., infrastructure failure and trapped citizens). It is imperative to achieve rapid identification of such events for timely and effective rainstorm disaster management. However, due to small-sample event classes, it becomes challenging to achieve the desired identification results. In this paper, a cross-domain transfer learning scheme is proposed to identify rainstorm events by combining citizen-report texts with multi-source spatial data. In the proposed framework, we recommend joint distribution adaptation (JDA) as the key to improving the identification of target events with inadequate samples by transferring knowledge from similar event domains. Meanwhile, multi-source spatial features are extracted and embedded in the text with linear discriminant analysis (LDA) to overcome the feature incompleteness that is caused by short text length. The proposed approach is verified based on the real data in Wuhan, China. Experimental results show that the knowledge transfer between different event classes helps cities to improve the performance of rainstorm event identification, with the accuracy increasing from 81% to 92%. Under data imbalance scenarios, the proposed scheme outperforms the state-of-the-art methods, such as sampling-based methods, ERNIE, and TCA. At the same time, the combination of textual and spatial features results in significant improvement in event identification, as evident by an 8% increase in accuracy. This framework can be followed to build a rainstorm event warning and assistance system to rapidly adapt management responses.

Transfer learning for multi-objective non-intrusive load monitoring in smart building

Buildings represent 39% of global greenhouse gas emissions, thus reducing carbon emissions in buildings is of importance to greenhouse gas emissions reductions. This requires understanding how electricity is utilized in the buildings, then optimizing electricity management to seek conservation of energy. Non-intrusive load monitoring (NILM) is a technique that disaggregates a house’s total load to estimate each appliance’s electric power usage. Several strategies for estimating one appliance at a time (one-to-one model) have been presented and experimentally proven to be effective, with two mainstream trends: appliance transfer learning and cross-domain transfer learning. The former refers to the transfer between different types of appliances in the same data domain, while the latter refers to the transfer between different data domains for the same type of appliance. Different from the previous work, this paper explores the approach of adopting one model for all appliances (one-to-many model) and proposes a novel transfer learning scheme, that incorporates appliance transfer learning and cross-domain transfer learning. Thus, a well-trained model can be transferred and utilized to effectively estimate the power consumption in another data set for all appliances, which demands fewer measurements and only one model. Three public data sets, REFIT, REDD, and UK-DALE, are used in our experiments. Further, a set of smart electricity meters was deployed in a practical non-residential building to validate the proposed method. The results demonstrate the accuracy and practicality compared to start-of-the-art one-to-one NILM transferred models.

Underwater crack pixel-wise identification and quantification for dams via lightweight semantic segmentation and transfer learning

Remotely operated vehicles (ROVs) with cameras provide a solution for dam underwater information acquisition, but problems like massive high-dimensional data processing and effective damage-related information extraction also occur. This paper thereby proposes a real-time pixel-level dam underwater crack automatic segmentation and quantification framework using lightweight semantic segmentation network LinkNet and two-stage hybrid transfer learning(TL). With the combination of in-domain and cross-domain TL, the modeling cost and computational burden can be significantly reduced by transferring knowledge learned in relevant domains to the target domain. The proposed method shows strong identification capability in complicated underwater scenarios(motion blur, uneven illumination, and obstacle blocking), achieving performance with 0.8924 mIOU, 0.9444 precision, 0.9151 recall, and 0.9295 F1-score in the test set. Combined with infrared laser-assisted ranging systems, the geometric features and physical sizes of cracks are quantified using the proposed method. Finally, a visual GUI software with both offline and online detection patterns is developed to perform real-time detection in practice.

An integrated underwater structural multi-defects automatic identification and quantification framework for hydraulic tunnel via machine vision and deep learning

Underwater structural defects in hydraulic tunnels are highly concealed and difficult to be identified by conventional manual methods. Remotely operated vehicle combined with visible light cameras can provide a noncontact and high spatial resolution damage detection solution. However, manually extracting useful structural damage-related information from massive data is time-consuming and involves high labor cost. This article proposes an integrated pixel-level underwater structural multi-defects instance segmentation and quantification framework for hydraulic tunnels via machine vision and deep learning. Firstly, a tunnel lining underwater structural multi-defects video dataset is developed. Next, an improved You Only Look At CoefficienTs for Edge devices is used to build the detector by exploiting temporal redundancy in videos. Three backbone detectors are used to trade off the balance between detection accuracy and efficiency, and a cross-domain transfer learning strategy is introduced to reduce model training costs and data dependencies. Various complicated tunnel underwater inspection scenarios, including uneven illumination, tilt shooting, high brightness, and motion blur scenarios, are used to evaluate model generalization capability. Experimental results show that ResNet50-based YolactEdge can well trade off the balance between accuracy and speed, which achieves 92.47 bbox mAP, 92.15 mask mAP, and 39.27 FPS in the testing set. A quantification evaluation method is proposed to quantify the detection results and extract the geometric features of structural defects based on digital image processing techniques. The proposed method can accurately identify the number, size, and area of tunnel underwater structural defects, providing data support for subsequent reinforcement.

Cross-domain intelligent bearing fault diagnosis under class imbalanced samples via transfer residual network augmented with explicit weight self-assignment strategy based on meta data

Intelligent fault diagnosis methods are significant to mitigate the dependency on expert knowledge and the cost. For the limited faulty data and variational working conditions of real scenarios, cross-domain diagnosis using existing diagnosis models is widely discussed. Especially, methods based on cross-domain transfer learning show great potentiality. However, the class imbalanced data of actual working conditions make it difficult to learn the actual fault feature distribution. To this end, a transfer residual network augmented with explicit weight self-assignment strategy based on meta data(TRN-EWM) is proposed. Specifically, we use a domain-shared ResNet to extract depth features of the data, which effectively avoid gradient disappearance and improve classification performance. Then, to lessen diagnosis difficulties in cross-domain and fully mine the actual feature distribution of the samples, a class imbalanced cross-domain transfer method is carried out. Ultimately, we creatively construct an explicit weight self-assignment strategy based on meta data for sample weight rebalancing, which prevents the dominance of major classes and the overfitting of minor classes. Two transfer experiments are conducted, and average cross-domain diagnosis accuracy of 99.60% is achieved by the proposed method, which shows the effectiveness in bearing fault diagnosis.

A novel bearing fault diagnosis method with feature selection and manifold embedded domain adaptation

Traditional fault diagnosis models assume that the training and test data sets have the same feature distribution, but in practice the distribution between the training and test sets varies considerably, making it difficult to achieve the desired fault diagnosis performance. Thus, a diagnosis method based on feature selection and manifold embedding domain adaptation is proposed in this paper. First, the signal is decomposed by variational modal decomposition to obtain multiple modal components, and the entropy, time domain and frequency domain features of each modal component are extracted to form a mixed feature set. Second, it proposes a feature evaluation index based on Fisher scores and feature domain differences to select features that are conducive to cross-domain fault diagnosis and transfer learning. Then, the geodesic flow core is constructed to learn the transformation feature representation in the Grassmann manifold space to avoid features are distorted. Finally, the domain adaptation is performed by minimizing the discrepancy in the joint probability distribution between the same category and maximizing the discrepancy between the different categories. Based on the results of multi-index experiments, the method in this paper is superior to other existing methods.

Cross-Domain Transfer Learning Prediction of COVID-19 Popular Topics Based on Knowledge Graph

The significance of research on public opinion monitoring of social network emergencies is becoming increasingly important. As a platform for users to communicate and share information online, social networks are often the source of public opinion about emergencies. Considering the relevance and transmissibility of the same event in different social networks, this paper takes the COVID-19 outbreak as the background and selects the platforms Weibo and TikTok as the research objects. In this paper, first, we use the transfer learning model to apply the knowledge obtained in the source domain of Weibo to the target domain of TikTok. From the perspective of text information, we propose an improved TC-LDA model to measure the similarity between the two domains, including temporal similarity and conceptual similarity, which effectively improves the learning effect of instance transfer and makes up for the problem of insufficient sample data in the target domain. Then, based on the results of transfer learning, we use the improved single-pass incremental clustering algorithm to discover and filter popular topics in streaming data of social networks. Finally, we build a topic knowledge graph using the Neo4j graph database and conduct experiments to predict the evolution of popular topics in new emergencies. Our research results can provide a reference for public opinion monitoring and early warning of emergencies in government departments.

Intra-domain and cross-domain transfer learning for time series data—How transferable are the features?

In practice, it is very challenging and sometimes impossible to collect datasets of labelled data large enough to successfully train a machine learning model, and one possible solution to this problem is using transfer learning. In this study, we investigate how transferable are features between different domains of time series data and under what conditions. The effects of transfer learning are observed in terms of the predictive performance of the models and their convergence rate during training. In our experiment, we used reduced datasets of 1500 and 9000 data instances to mimic real-world conditions. We trained two sets of models (four different architectures) on the reduced datasets: those trained with transfer learning and those trained from scratch. Knowledge transfer was performed both within the same application domain (seismology) and between different application domains (seismology, speech, medicine, finance). We observed the prediction performance of the models and their training convergence rate. We repeated the experiments seven times and applied statistical tests to confirm the validity of the results. The overall conclusion of our study is that transfer learning is highly likely to either increase or not negatively affect the model’s predictive performance or its training convergence rate. We discuss which source and target domains are compatible for knowledge transfer. We also discuss the effect of the target dataset size and the choice of the model and its hyperparameters on transfer learning.

Deep Transfer Learning for Biology Cross-Domain Image Classification

Automatic biology image classification is essential for biodiversity conservation and ecological study. Recently, due to the record-shattering performance, deep convolutional neural networks (DCNNs) have been used more often in biology image classification. However, training DCNNs requires a large amount of labeled data, which may be difficult to collect for some organisms. This study was carried out to exploit cross-domain transfer learning for DCNNs with limited data. According to the literature, previous studies mainly focus on transferring from ImageNet to a specific domain or transferring between two closely related domains. While this study explores deep transfer learning between species from different domains and analyzes the situation when there is a huge difference between the source domain and the target domain. Inspired by the analysis of previous studies, the effect of biology cross-domain image classification in transfer learning is proposed. In this work, the multiple transfer learning scheme is designed to exploit deep transfer learning on several biology image datasets from different domains. There may be a huge difference between the source domain and the target domain, causing poor performance on transfer learning. To address this problem, multistage transfer learning is proposed by introducing an intermediate domain. The experimental results show the effectiveness of cross-domain transfer learning and the importance of data amount and validate the potential of multistage transfer learning.

Tri-training for Dependency Parsing Domain Adaptation

In recent years, the research on dependency parsing focuses on improving the accuracy of the domain-specific (in-domain) test datasets and has made remarkable progress. However, there are innumerable scenarios in the real world that are not covered by the dataset, namely, the out-of-domain dataset. As a result, parsers that perform well on the in-domain data usually suffer from significant performance degradation on the out-of-domain data. Therefore, to adapt the existing in-domain parsers with high performance to a new domain scenario, cross-domain transfer learning methods are essential to solve the domain problem in parsing. This paper examines two scenarios for cross-domain transfer learning: semi-supervised and unsupervised cross-domain transfer learning. Specifically, we adopt a pre-trained language model BERT for training on the source domain (in-domain) data at the subword level and introduce self-training methods varied from tri-training for these two scenarios. The evaluation results on the NLPCC-2019 shared task and universal dependency parsing task indicate the effectiveness of the adopted approaches on cross-domain transfer learning and show the potential of self-learning to cross-lingual transfer learning.

Transfer learning method for rolling bearing fault diagnosis under different working conditions based on CycleGAN

In practical bearing fault diagnosis tasks, the available labeled data are often not from the equipment to be diagnosed and cannot cover all manner of working conditions. The adopted data-driven method is required to have a certain degree of cross-domain and cross-working condition transfer learning diagnosis ability. However, limited by the performance of existing transfer learning methods, the potential difference between the source domain and the target domain poses a challenge for the accuracy of transfer diagnosis. In this paper, a cross-working condition data supplement method based on the cycle generative adversarial network (CycleGAN) and a dynamics model is proposed, which can use limited available data to approximate the missing parts of existing data and be used for diagnosis of the target domain. First, we considered the limited experimental data as the target domain, the simulation data corresponding to the working condition as the source domain and used the working condition as the benchmark to constrain the data correspondence between the two datasets. We then used the CycleGAN model to learn the feature mapping from simulation to experiment. Second, based on the working condition of the data to be tested, the corresponding simulation data were input into the trained generator to obtain labeled data with experimental characteristics under the corresponding working conditions, and transferred the dataset as the source domain data to the data to be tested. In the test using self-made simulation and experimental datasets, combined with the transfer learning method based on the probability distribution adaptation, it was shown that the proposed method could effectively improve the diagnostic impact of the single transfer learning method in cross-domain and cross-working conditions when the working condition span was large.

COVID-19 lung infection segmentation with a novel two-stage cross-domain transfer learning framework.

With the global outbreak of COVID-19 in early 2020, rapid diagnosis of COVID-19 has become the urgent need to control the spread of the epidemic. In clinical settings, lung infection segmentation from computed tomography (CT) images can provide vital information for the quantification and diagnosis of COVID-19. However, accurate infection segmentation is a challenging task due to (i) the low boundary contrast between infections and the surroundings, (ii) large variations of infection regions, and, most importantly, (iii) the shortage of large-scale annotated data. To address these issues, we propose a novel two-stage cross-domain transfer learning framework for the accurate segmentation of COVID-19 lung infections from CT images. Our framework consists of two major technical innovations, including an effective infection segmentation deep learning model, called nCoVSegNet, and a novel two-stage transfer learning strategy. Specifically, our nCoVSegNetconducts effective infection segmentation by taking advantage of attention-aware feature fusion and large receptive fields, aiming to resolve the issues related to low boundary contrast and large infection variations. To alleviate the shortage of the data, the nCoVSegNetis pre-trained using a two-stage cross-domain transfer learning strategy, which makes full use of the knowledge from natural images (i.e., ImageNet) and medical images (i.e., LIDC-IDRI) to boost the final training on CT images with COVID-19 infections. Extensive experiments demonstrate that our framework achieves superior segmentation accuracy and outperforms the cutting-edge models, both quantitatively and qualitatively.