Low-resolution Feature Map Research Articles

Lightweight medical image segmentation network with multi-scale feature-guided fusion

In the field of computer-aided medical diagnosis, it is crucial to adapt medical image segmentation to limited computing resources. There is tremendous value in developing accurate, real-time vision processing models that require minimal computational resources. When building lightweight models, there is always a trade-off between computational cost and segmentation performance. Performance often suffers when applying models to meet resource-constrained scenarios characterized by computation, memory, or storage constraints. This remains an ongoing challenge. This paper proposes a lightweight network for medical image segmentation. It introduces a lightweight transformer, proposes a simplified core feature extraction network to capture more semantic information, and builds a multi-scale feature interaction guidance framework. The fusion module embedded in this framework is designed to address spatial and channel complexities. Through the multi-scale feature interaction guidance framework and fusion module, the proposed network achieves robust semantic information extraction from low-resolution feature maps and rich spatial information retrieval from high-resolution feature maps while ensuring segmentation performance. This significantly reduces the parameter requirements for maintaining deep features within the network, resulting in faster inference and reduced floating-point operations (FLOPs) and parameter counts. Experimental results on ISIC2017 and ISIC2018 datasets confirm the effectiveness of the proposed network in medical image segmentation tasks. For instance, on the ISIC2017 dataset, the proposed network achieved a segmentation accuracy of 82.33 % mIoU, and a speed of 71.26 FPS on 256 × 256 images using a GeForce GTX 3090 GPU. Furthermore, the proposed network is tremendously lightweight, containing only 0.524M parameters. The corresponding source codes are available at https://github.com/CurbUni/LMIS-lightweight-network.

An enhanced approach for few-shot segmentation via smooth downsampling mask and label smoothing loss

Few-shot semantic segmentation aims to segment new categories with only a small number of annotated images. Previous methods mainly focused on exploiting the pixel-level correlation between the support image and the query image, combined with attention-based methods, resulting in significant advancements. In this paper, we introduce a new perspective to enhance few-shot segmentation. We identify that utilizing the bilinear interpolation method to downsample the mask leads to the loss of fine-grained information from the target features. To address this issue, we propose a Smooth Downsampling Mask (SDM) method. The SDM method is designed to retain more effective target semantic features by employing a cascaded downsampling approach with a smooth kernel for mask processing. Additionally, we propose a label smoothing loss to further enhance the performance, which provides direct guidance for low-resolution feature map optimization. Both methods can be used as plug-and-play modules for existing methods. Notably, our proposed method does not involve additional learnable parameters and is computationally efficient, thus achieving painless gains. To validate the effectiveness of our method, we take three publicly available models as baselines and conduct extensive experiments on three public benchmarks PASCAL-5i, COCO-20i and FSS-1000, and achieve considerable improvement.

Attentional decoder networks for chest X-ray image recognition on high-resolution features

Background and objective:This paper introduces an encoder–decoder-based attentional decoder network to recognize small-size lesions in chest X-ray images. In the encoder-only network, small-size lesions disappear during the down-sampling steps or are indistinguishable in the low-resolution feature maps. To address these issues, the proposed network processes images in the encoder–decoder architecture similar to U-Net families and classifies lesions by globally pooling high-resolution feature maps. However, two challenging obstacles prohibit U-Net families from being extended to classification: (1) the up-sampling procedure consumes considerable resources, and (2) there needs to be an effective pooling approach for the high-resolution feature maps. Methods:Therefore, the proposed network employs a lightweight attentional decoder and harmonic magnitude transform. The attentional decoder up-samples the given features with the low-resolution features as the key and value while the high-resolution features as the query. Since multi-scaled features interact, up-sampled features embody global context at a high resolution, maintaining pathological locality. In addition, harmonic magnitude transform is devised for pooling high-resolution feature maps in the frequency domain. We borrow the shift theorem of the Fourier transform to preserve the translation invariant property and further reduce the parameters of the pooling layer by an efficient embedding strategy. Results:The proposed network achieves state-of-the-art classification performance on the three public chest X-ray datasets, such as NIH, CheXpert, and MIMIC-CXR. Conclusions:In conclusion, the proposed efficient encoder–decoder network recognizes small-size lesions well in chest X-ray images by efficiently up-sampling feature maps through an attentional decoder and processing high-resolution feature maps with harmonic magnitude transform. We open-source our implementation at https://github.com/Lab-LVM/ADNet.

Efficient Small-Object Detection in Underwater Images Using the Enhanced YOLOv8 Network

Underwater object detection plays a significant role in marine ecosystem research and marine species conservation. The improvement of related technologies holds practical significance. Although existing object-detection algorithms have achieved an excellent performance on land, they are not satisfactory in underwater scenarios due to two limitations: the underwater objects are often small, densely distributed, and prone to occlusion characteristics, and underwater embedded devices have limited storage and computational capabilities. In this paper, we propose a high-precision, lightweight underwater detector specifically optimizing for underwater scenarios based on the You Only Look Once Version 8 (YOLOv8) model. Firstly, we replace the Darknet-53 backbone of YOLOv8s with FasterNet-T0, reducing model parameters by 22.52%, FLOPS by 23.59%, and model size by 22.73%, achieving model lightweighting. Secondly, we add a Prediction Head for Small Objects, increase the number of channels for high-resolution feature map detection heads, and decrease the number of channels for low-resolution feature map detection heads. This results in a 1.2% improvement in small-object detection accuracy, while the remaining model parameters and memory consumption are nearly unchanged. Thirdly, we use Deformable ConvNets and Coordinate Attention in the neck part to enhance the accuracy in the detection of irregularly shaped and densely occluded small targets. This is achieved by learning convolution offsets from feature maps and emphasizing the regions of interest (RoIs). Our method achieves 52.12% AP on the underwater dataset UTDAC2020, with only 8.5 M parameters, 25.5 B FLOPS, and 17 MB model size. It surpasses the performance of large model YOLOv8l, at 51.69% AP, with 43.6 M parameters, 164.8 B FLOPS, and 84 MB model size. Furthermore, by increasing the input image resolution to 1280 × 1280 pixels, our model achieves 53.18% AP, making it the state-of-the-art (SOTA) model for the UTDAC2020 underwater dataset. Additionally, we achieve 84.4% mAP on the Pascal VOC dataset, with a substantial reduction in model parameters compared to previous, well-established detectors. The experimental results demonstrate that our proposed lightweight method retains effectiveness on underwater datasets and can be generalized to common datasets.

Gradient-guided hierarchical feature attack for object detector

Deep neural networks (DNNs) are vulnerable to adversarial attacks, which can cause security risks in computer information systems. Feature disruption attacks, as a typical form of adversarial attack, optimize adversarial examples by disrupting the intermediate features extracted by DNN. The existing feature disruption attacks have limitations when it comes to objects of different scales within resolution features, favoring low-resolution feature maps and low-scoring objects. The imbalance above affects their effectiveness in object detection tasks. Gradient-guided Hierarchical Feature Attack (GHFA) is proposed to solve these problems, which disrupts detector-extracted features using gradient-guided feature weighting. GHFA incorporates receptive field scaling and gradient scaling to balance the focus among feature maps and enhance attack performance. The evaluation of GHFA on 9 object detectors demonstrates its superior transferability compared with the 6 comparative methods, surpassing the second-ranked method by 2.4%. Furthermore, the real-world applicability of GHFA is validated available by testing it on a commercial online object detection platform.

Coarse-to-fine matching via cross fusion of satellite images

The registration of multimodal satellite images is essential for a prerequisite for accruing complementary observational data. Nevertheless, the differential imaging nuances amongst non-linear radiometric multimodal images precipitate a complexity in keypoint detection, rendering it a great challenge. This complexity exacerbates the difficulty encountered in matching multimodal satellite images. In this paper, a dual-branch cross fusion network (DF-Net) is proposed for the purpose of satellite image registration. DF-Net relies on the self-attention granted to a pair of images, thereby providing cross-modal fusion feature descriptions. Initially, reference and sensed images are deployed as inputs for the dual-branch network, which in turn engenders feature descriptions of both high and low resolution, respectively. Sequentially, the matching of individual feature descriptions is anchored on the low-resolution feature map, paving the way for the establishment of coarse matching correspondences. Subsequently, the outcomes of these coarse correspondences are transposed onto the feature map with a higher resolution, thereby generating fine matching results for each coarse correspondence. An exhaustive set of qualitative and quantitative assessments have been administered on three satellite image datasets encompassing a diverse range of scenarios. The average Repeatability (Rep.), Mean Matching Accuracy (MMA), and Root-Mean-Square Error (RMSE) of the DF-Net applied to three large-scale satellite images were recorded to be 0.71, 0.65, and 2.34, respectively. These findings buttress the proficiency of the proposed strategy in facilitating cross-modal matching and bear testimony to the sterling performance of the method proposed.

HS-YOLO: Small Object Detection for Power Operation Scenarios

Object detection methods are commonly employed in power safety monitoring systems to detect violations in surveillance scenes. However, traditional object detection methods are ineffective for small objects that are similar to the background information in the power monitoring scene, which consequently affects the performance of violation behavior detection. This paper proposed a small object detection algorithm named HS-YOLO, based on High-Resolution Network (HRNet) and sub-pixel convolution. First, to fully extract the microfeature information of the object, a small object feature extraction backbone network is proposed based on the HRNet structure. The feature maps of different scales are processed by multiple parallel branches and fused with each other in the network. Then, to fully retain the effective features of small objects, the sub-pixel convolution module is incorporated as the upsampling operator in the feature fusion network. The low-resolution feature map is upsampled to a higher resolution by reorganizing pixel values and performing padding operations in this module. On our self-constructed power operation dataset, the HS-YOLO algorithm achieved a mAP of 87.2%, which is a 3.5% improvement compared to YOLOv5. Particularly, the dataset’s AP for detecting small objects such as cuffs, necklines, and safety belts is improved by 10.7%, 5.8%, and 4.4%, respectively. These results demonstrate the effectiveness of our proposed method in detecting small objects in power operation scenarios.

SADENet: A supervised attention delicate enhanced network for subtle person detection

Object detection algorithms have been developed for decades, and one of remaining important challenges is the detection of subtle targets in complex environments. Many convolution neural network (CNN) based methods have been proposed to handle this problem. However, the information pertaining to subtle targets is delicate and can easily be lost or neglected in the upsampling and downsampling stages of conventional CNN. In this paper, we propose a new subtle target detection network, named the supervised attention delicate enhanced network (SADENet). Firstly, two sample vertebras are designed to solve the problem of information loss during upsampling and downsampling. Then, an enhanced context attention module (ECAM) is proposed, which extends the original supervised heatmap by fusing information from different scales. Based on the ECAM and the respective advantages of high-resolution and low-resolution feature maps, two effective sampling attention modules, the semantic fidelity upsampling module and the location enhancement downsampling module, are created to build a bi-directional connection neck structure. The effectiveness of our proposed method is demonstrated in experiments on the TinyPerson dataset, which outperforms many state-of-the-art detectors, especially in the detection of tiny and tiny1 scale targets.

SSP-Net: Scalable sequential pyramid networks for real-Time 3D human pose regression

In this paper we propose a highly scalable convolutional neural networks, end-to-end trainable, for real-time 3D human pose regression from still RGB images. We call this approach Scalable Sequential Pyramid Networks (SSP-Net) as it is trained with refined supervision at multiple scales in a sequential manner. Our network requires a single training procedure and is capable of producing its best predictions at 120 frames per second (FPS), or acceptable predictions at more than 200 FPS when cut at test time. We show that the proposed regression approach is invariant to the size of feature maps, allowing our method to perform multi-resolution intermediate supervisions and reaching results comparable to the state-of-the-art with very low resolution feature maps. We demonstrate the accuracy and the effectiveness of our method by providing extensive experiments on two of the most important publicly available datasets for 3D pose estimation, Human3.6M and MPI-INF-3DHP. Additionally, we provide relevant insights about our decisions on the network architecture and show its flexibility to meet the best precision-speed compromise.

High-Resolution Swin Transformer for Automatic Medical Image Segmentation.

The resolution of feature maps is a critical factor for accurate medical image segmentation. Most of the existing Transformer-based networks for medical image segmentation adopt a U-Net-like architecture, which contains an encoder that converts the high-resolution input image into low-resolution feature maps using a sequence of Transformer blocks and a decoder that gradually generates high-resolution representations from low-resolution feature maps. However, the procedure of recovering high-resolution representations from low-resolution representations may harm the spatial precision of the generated segmentation masks. Unlike previous studies, in this study, we utilized the high-resolution network (HRNet) design style by replacing the convolutional layers with Transformer blocks, continuously exchanging feature map information with different resolutions generated by the Transformer blocks. The proposed Transformer-based network is named the high-resolution Swin Transformer network (HRSTNet). Extensive experiments demonstrated that the HRSTNet can achieve performance comparable with that of the state-of-the-art Transformer-based U-Net-like architecture on the 2021 Brain Tumor Segmentation dataset, the Medical Segmentation Decathlon's liver dataset, and the BTCV multi-organ segmentation dataset.

Progressive Transformer Machine for Natural Character Reenactment

Character reenactment aims to control a target person’s full-head movement by a driving monocular sequence that is made up of the driving character video. Current algorithms utilize convolution neural networks in generative adversarial networks, which extract historical and geometric information to iteratively generate video frames. However, convolution neural networks can merely capture local information with limited receptive fields and ignore global dependencies that play a crucial role in face synthesis, leading to generating unnatural video frames. In this work, we design a progressive transformer module that introduces multi-head self-attention with convolution refinement to simultaneously capture global-local dependencies. Specifically, we utilize the non-lapping window-based multi-head self-attention mechanism with hierarchical architecture to obtain the larger receptive fields at low-resolution feature map and thus extract global information. To better model local dependencies, we introduce the convolution operation to further refine the attentional weight in the multi-head self-attention mechanism. Finally, we use several stacked progressive transformer modules with the down-sampling operation to encode information of appearance information of previously generated frames and parameterized 3D face information of the current frame. Similarly, we use several stacked progressive transformer modules with the up-sampling operation to iteratively generate video frames. In this way, it can capture global-local information to facilitate generating video frames that are globally natural while preserving sharp outlines and rich detail information. Extensive experiments on several standard benchmarks suggest that the proposed method outperforms current leading algorithms.

Fast-HBNet: Hybrid Branch Network for Fast Lane Detection

As one of the fundamental visual tasks in the unmanned driving area, lane detection attracts increasing attention. In practical applications, lane points are very difficult to localize because they usually appear to be sparse and incomplete due to the influence of illumination and environment. Conventional lane detection methods rely on coarse features and carefully designed postprocessing to detect the lane lines. However, these methods are usually slow, and the stability and generalization ability are unsatisfactory. In this paper, we propose a novel lane detection network – Fast-HBNet(Fast-Hybrid Branch Network), which exploits both global semantic information and spatial contexts. To enlarge receptive fields and encode more detailed information, the compound transformation is employed and the proposed hybrid branch network extracts four diverse feature maps with different receptive fields and spatial contexts. Besides, we design a Hierarchical Feature Learning (HFL) module to learn lane features from the scale, channel, and spatial levels to enhance the generalization ability of our detector. These features are further selectively coalesced to generate unified lane feature maps with large receptive fields and rich detailed information. In other words, our network can encode the global semantic information from the high-resolution feature maps and the fine-grained details in the low-resolution feature maps. Experimental results conducted on the TuSimple (2017) and CULane [Pan <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">et al.</i> (2018)] datasets demonstrate that the proposed Fast-HBNet outperforms numerous state-of-the-art lane detectors in both speed and accuracy. Particularly, Fast-HBNet achieves an accuracy of 96.88% on the TuSimple dataset at a speed of 76 FPS.

Balanced Spatial Feature Distillation and Pyramid Attention Network for Lightweight Image Super-resolution

Recently, the attention mechanism became the key issue for image super-resolution (SR) because it has the ability to extract different features from the image according to the used attention type. Despite the great success of attention-based methods, the conflict among features of different attention types can affect the SR performance. In this paper, we propose an efficient single image SR model called balanced spatial feature distillation and pyramid attention (BSPAN). The idea of BSPAN is based on the trade-off among the extracted features of different attention types. Also, we propose a balanced spatial feature distillation block (BSFDB) as the backbone of BSPAN so that the network can effectively advantage from the different attention features. Two different attention types, namely spatial attention residual feature distillation (SARFD) and classical attention (CA) are considered in the BSFDB to achieve a leveling between them based on the contents of the low-resolution feature map using the balancing attention block. The BSFDB block is designed to improve the SR performance and has lightweight parameters and low computation complexity. Moreover, to further improve the SR performance, the pyramid attention is introduced in the middle of the BSPAN network for extracting long-range features at a variety of locations and scales. Evaluation based on five benchmark datasets, we concluded that the balancing between features of a variety of attention types can effectively improve the SR performance. So, the proposed BSPAN model achieves significant enhancements in comparison with the state-of-the-art and superior visual quality and reconstruction accuracy.

Deep feature pyramid network for EEG emotion recognition

In this work, the Feature Pyramid Network (FPN) is proposed for improving the performance of electroencephalography (EEG) emotion recognition. Differential Entropy (DE) is extracted from each EEG channel as the basic feature. Then the feature matrix was constructed through biharmonic spline interpolation to obtain the correlation information between EEG channels. Next, the proposed FPN is applied to fuse multiscale EEG spatial information for obtaining deeper semantic features. FPN integrates the semantic information of low-level and high-level feature maps through up-sampling and lateral connections. It can reduce the semantic gap between feature maps and increase the receptive field of low-resolution feature maps. Finally, a linear Support Vector Machine (SVM) is utilized for emotion recognition. The experiment results of the proposed method have a state-of-the-art performance of valence (94.29%) and arousal (96.97%) by 70–30 cross-validation on the Database for Emotion Analysis using Physiological Signals (DEAP). And the accuracy of 97.12% was obtained on the SJTU Emotion EEG Dataset (SEED). Moreover, we carried out the LOSO CV experiment on DEAP. The accuracy of valence and arousal reached 80.38% and 82.33%, respectively.

Multi-Scale Semi-Coupled Convolutional Sparse Coding for the Super-Resolution Reconstruction of Remote Sensing Image

The traditional convolutional sparse coding super-resolution method only introduces the linear projection relationship in the feature space conversion and fails to consider the local detail in the feature map learning, which causes the reconstruction results to be unsatisfactory in terms of edges and details. The convolutional sparse coding theory is introduced into the super-resolution reconstruction framework of remote sensing images and a multi-scale semi-coupled convolutional sparse coding super-resolution reconstruction method is proposed. Firstly, the input image is decomposed by multi-scale to extract smooth components and multi-scale texture components, and the final smoothing components are reconstructed by bicubic interpolation. Then, the texture components of each scale are reconstructed by semi-coupled convolution sparse coding. The nonlinear convolution operator is used as the projection function between the high-resolution feature map and the low-resolution feature map of texture component at each scale and the non-local self-similarity structure in the feature map learning for constrained optimization is introduced to better reconstruct the texture image at each scale. Finally, the reconstructed smooth component and the texture component at each scale are superimposed to obtain the final reconstructed image. Remote sensing images from 4 different sensors are used as experimental images and the state-of-the-art super-resolution reconstruction methods are compared. Experimental results show that the reconstructed images obtained by the proposed method are better than other methods in quantitative analysis index PSNR and FSIM and show clearer boundary and detailed information and have a certain anti-noise performance.

Concrete crack segmentation based on UAV-enabled edge computing

In recent years, the rapid development of UAV technology has greatly improved the efficiency of the detection of concrete bridge cracks. With the increase in the number of bridge inspection UAVs, the number of tasks handled by cloud services has increased linearly, resulting in increased computational pressure on cloud services. In order to reduce the computational load of cloud servers, we proposed a crack segmentation network based on UAV-enabled edge computing. However, due to the limitation of computational capability of edge computing and the strength inhomogeneity and background complexity of cracks, crack detection is still a challenging task. Thus, we proposed an effective concrete crack segmentation network based on UAV-enabled edge computing, the network used feature map fusion to fuse different levels of feature map information into lower-level features for crack detection. The atrous spatial pyramid pooling network was used to increase the low-resolution feature map receptive field information for cracks and to enhance the detection accuracy for cracks of different scales. In addition, loss functions for crack datasets were proposed to solve the problem of imbalance due to positive and negative samples in the concrete crack images. Experiments demonstrated that the proposed methods are better than the state-of-the-art edge detection and semantic segmentation methods in terms of accuracy and generality.

Multi-attention augmented network for single image super-resolution

How to improve the representational power of visual features extracted by deep convolutional neural networks is of crucial importance for high-quality image super-resolution. To address this issue, we propose a multi-attention augmented network, which mainly consists of content-, orientation- and position-aware modules. Specifically, we develop an attention augmented U-net structure to form the content-aware module in order to learn and combine multi-scale informative features within a large receptive field. To better reconstruct image details in different directions, we design a set of pre-defined sparse kernels to construct the orientation-aware module, which can extract more representative multi-orientation features and enhance the discriminative capacity in stacked convolutional stages. Then these extracted features are adaptively fused through channel attention mechanism. In upscale stage, the position-aware module adopts a novel self-attention to reweight the element-wise value of final low-resolution feature maps, for further suppressing the possible artifacts. Experimental results demonstrate that our method obtains better reconstruction accuracy and perceptual quality against state-of-the-art methods.

Pixel-in-Pixel Net: Towards Efficient Facial Landmark Detection in the Wild

Recently, heatmap regression models have become popular due to their superior performance in locating facial landmarks. However, three major problems still exist among these models: (1) they are computationally expensive; (2) they usually lack explicit constraints on global shapes; (3) domain gaps are commonly present. To address these problems, we propose Pixel-in-Pixel Net (PIPNet) for facial landmark detection. The proposed model is equipped with a novel detection head based on heatmap regression, which conducts score and offset predictions simultaneously on low-resolution feature maps. By doing so, repeated upsampling layers are no longer necessary, enabling the inference time to be largely reduced without sacrificing model accuracy. Besides, a simple but effective neighbor regression module is proposed to enforce local constraints by fusing predictions from neighboring landmarks, which enhances the robustness of the new detection head. To further improve the cross-domain generalization capability of PIPNet, we propose self-training with curriculum. This training strategy is able to mine more reliable pseudo-labels from unlabeled data across domains by starting with an easier task, then gradually increasing the difficulty to provide more precise labels. Extensive experiments demonstrate the superiority of PIPNet, which obtains state-of-the-art results on three out of six popular benchmarks under the supervised setting. The results on two cross-domain test sets are also consistently improved compared to the baselines. Notably, our lightweight version of PIPNet runs at 35.7 FPS and 200 FPS on CPU and GPU, respectively, while still maintaining a competitive accuracy to state-of-the-art methods. The code of PIPNet is available at https://github.com/jhb86253817/PIPNet.

Gradual back-projection residual attention network for magnetic resonance image super-resolution

Background and objectiveMagnetic Resonance Image (MRI) analysis can provide anatomical examination of internal organs, which is helpful for diagnosis of the disease. Aiming at the problems of insufficient feature information mining in the process of MRI super-resolution (SR) reconstruction, the difficulty of determining the interdependence between the channels of the feature map, and the reconstruction error when reconstructing high-resolution (HR) images, we propose a SR method to solve these problems. MethodsIn this work, we propose a gradual back-projection residual attention network for MRI super-resolution (GRAN), which outperforms most of the state-of-the-art methods. Firstly, we use the gradual upsampling method to gradually scale the low-resolution (LR) image to a given magnification to alleviate the high-frequency information loss caused by the upsampling process. Secondly, we merge the idea of iterative back-projection at each stage of gradual upsampling, learn the mapping relationship between HR and LR feature maps and reduce the noise introduced during the upsampling process. Finally, we use the attention mechanism to dynamically allocate attention resources to the feature maps generated at different stages of the gradual back-projection network, so that the network model can learn the interdependence between each feature map. ResultsFor the 2 × and 4 × enlargement, the proposed GRAN method shows the superiority over the state-of-the-art methods on the Set5, Set14, and Urban100 benchmark datasets, extensive benchmark experiment and analysis show that the superiority of the GRAN algorithm in terms of peak signal-to-noise ratio and structural similarity index indicators. ConclusionThe MRI results reconstructed by gradual back-projection residual attention network on the public dataset IDI have good image sharpness, rich texture details and good visual experience. In addition, the reconstructed image is the closest to the real image, enabling the medical expert to see the biological tissue structure and its early pathological changes more clearly, providing assistance and support to the medical expert in the diagnosis and treatment of the disease.

3D axial-attention for lung nodule classification

In recent years, Non-Local-based methods have been successfully applied to lung nodule classification. However, these methods offer 2D attention or limited 3D attention to low-resolution feature maps. Moreover, they still depend on a convenient local filter such as convolution as full 3D attention is expensive to compute and requires a big dataset, which might not be available. We propose to use 3D Axial-Attention, which requires a fraction of the computing power of a regular Non-Local network (i.e., self-attention). Unlike a regular Non-Local network, the 3D Axial-Attention network applies the attention operation to each axis separately. Additionally, we solve the invariant position problem of the Non-Local network by proposing to add 3D positional encoding to shared embeddings. We validated the proposed method on 442 benign nodules and 406 malignant nodules, extracted from the public LIDC-IDRI dataset by following a rigorous experimental setup using only nodules annotated by at least three radiologists. Our results show that the 3D Axial-Attention model achieves state-of-the-art performance on all evaluation metrics, including AUC and Accuracy. The proposed model provides full 3D attention, whereby every element (i.e., pixel) in the 3D volume space attends to every other element in the nodule effectively. Thus, the 3D Axial-Attention network can be used in all layers without the need for local filters. The experimental results show the importance of full 3D attention for classifying lung nodules.