Tissue self-attention network for the segmentation of optical coherence tomography images on the esophagus

Automatic segmentation of layered tissue is the key to esophageal optical coherence tomography (OCT) image processing. With the advent of deep learning techniques, frameworks based on a fully convolutional network are proved to be effective in classifying pixels on images. However, due to speckle noise and unfavorable imaging conditions, the esophageal tissue relevant to the diagnosis is not always easy to identify. An effective approach to address this problem is extracting more powerful feature maps, which have similar expressions for pixels in the same tissue and show discriminability from those from different tissues. In this study, we proposed a novel framework, called the tissue self-attention network (TSA-Net), which introduces the self-attention mechanism for esophageal OCT image segmentation. The self-attention module in the network is able to capture long-range context dependencies from the image and analyzes the input image in a global view, which helps to cluster pixels in the same tissue and reveal differences of different layers, thus achieving more powerful feature maps for segmentation. Experiments have visually illustrated the effectiveness of the self-attention map, and its advantages over other deep networks were also discussed.     

Retinal optical coherence tomography image analysis by a restricted Boltzmann machine

Optical coherence tomography (OCT) is an emerging imaging technique for ophthalmic disease diagnosis. Two major problems in OCT image analysis are image enhancement and image segmentation. Deep learning methods have achieved excellent performance in image analysis. However, most of the deep learning-based image analysis models are supervised learning-based approaches and need a high volume of training data (e.g., reference clean images for image enhancement and accurate annotated images for segmentation). Moreover, acquiring reference clean images for OCT image enhancement and accurate annotation of the high volume of OCT images for segmentation is hard. So, it is difficult to extend these deep learning methods to the OCT image analysis. We propose an unsupervised learning-based approach for OCT image enhancement and abnormality segmentation, where the model can be trained without reference images. The image is reconstructed by Restricted Boltzmann Machine (RBM) by defining a target function and minimizing it. For OCT image enhancement, each image is independently learned by the RBM network and is eventually reconstructed. In the reconstruction phase, we use the ReLu function instead of the Sigmoid function. Reconstruction of images given by the RBM network leads to improved image contrast in comparison to other competitive methods in terms of contrast to noise ratio (CNR). For anomaly detection, hyper-reflective foci (HF) as one of the first signs in retinal OCTs of patients with diabetic macular edema (DME) are identified based on image reconstruction by RBM and post-processing by removing the HFs candidates outside the area between the first and the last retinal layers. Our anomaly detection method achieves a high ability to detect abnormalities.     

Deep weakly-supervised breast tumor segmentation in ultrasound images with explicit anatomical constraints

Breast tumor segmentation is an important step in the diagnostic procedure of physicians and computer-aided diagnosis systems. We propose a two-step deep learning framework for breast tumor segmentation in breast ultrasound (BUS) images which requires only a few manual labels. The first step is breast anatomy decomposition handled by a semi-supervised semantic segmentation technique. The input BUS image is decomposed into four breast anatomical structures, namely fat, mammary gland, muscle and thorax layers. Fat and mammary gland layers are used as constrained region to reduce the search space for breast tumor segmentation. The second step is breast tumor segmentation performed in a weakly-supervised learning scenario where only image-level labels are available. Breast tumors are first recognized by a classification network and then segmented by the proposed class activation mapping and deep level set (CAM-DLS) method. For breast anatomy decomposition, the proposed framework achieves Dice similarity coefficient (DSC) of 83.0 ± 11.8%, 84.3 ± 10.0%, 80.7 ± 15.4% and 91.0 ± 11.4% for fat, mammary gland, muscle and thorax layers, respectively. For breast tumor recognition, the proposed framework achieves sensitivity of 95.8%, precision of 92.4%, specificity of 93.9%, accuracy of 94.8% and F1-score of 0.941. For breast tumor segmentation, the proposed framework achieves DSC of 77.3% and intersection-over-union (IoU) of 66.0%. In conclusion, the proposed framework could efficiently perform breast tumor recognition and segmentation simultaneously in a weakly-supervised setting with anatomical constraints.     

深度学习在医学影像中的应用

医学影像是放射科医生做出医学诊断的重要依据。但随着医学影像技术的快速发展, 逐渐增多的影像图像和复杂的图像信息对医生的工作产生了巨大的挑战。而深度学习是人工智能研究中最热门的领域, 在处理大数据和提取有效信息方面具有优势, 因此逐渐成为分析医学影像方面的首选方法。本文阐述了深度学习的概念, 并简要总结深度学习在医学影像中的常见模型, 包括卷积神经网络、循环神经网络、深度置信网络和自动编码器。卷积神经网络的基本结构是卷积层、池化层和全连接层; 循环神经网络由输入层、隐藏层和输出层组成; 深度置信网络的基础是玻尔兹曼机; 自动编码器包含编码层、隐藏层和解码层。通过对CT肺结节和MRI脑部疾病的分类, 阐明目前深度学习在疾病自动分类上准确性较高; 通过分割左心室、椎旁肌肉和肝脏的结构, 可见深度学习方法在医学图像分割上与人为分割具有一致性; 深度学习在肺结节和乳腺癌疾病的检测上已相对成熟。但目前为止, 仍存在标注的样本量少和过拟合的问题, 希望通过共享图像数据库来解决此问题。总之, 深度学习在医学影像中具有广阔前景, 且对临床医生的工作具有重大意义。      

ResGANet: Residual group attention network for medical image classification and segmentation

In recent years, deep learning technology has shown superior performance in different fields of medical image analysis. Some deep learning architectures have been proposed and used for computational pathology classification, segmentation, and detection tasks. Due to their simple, modular structure, most downstream applications still use ResNet and its variants as the backbone network. This paper proposes a modular group attention block that can capture feature dependencies in medical images in two independent dimensions: channel and space. By stacking these group attention blocks in ResNet-style, we obtain a new ResNet variant called ResGANet. The stacked ResGANet architecture has 1.51–3.47 times fewer parameters than the original ResNet and can be directly used for downstream medical image segmentation tasks. Many experiments show that the proposed ResGANet is superior to state-of-the-art backbone models in medical image classification tasks. Applying it to different segmentation networks can improve the baseline model in medical image segmentation tasks without changing the network architecture. We hope that this work provides a promising method for enhancing the feature representation of convolutional neural networks (CNNs) in the future.     

基于级联结构的肺癌患者颈部淋巴结超声图像分割研究

为了正确诊断肺癌转移，本文应用深度学习技术对肺癌患者颈部淋巴结超声图像病灶区域分割，提出了一种用于超声图像分割的级联注意力UNet网络，该级联结构是将注意力UNet与EfficientNet相结合的二阶段分割网络，第一阶段为粗分割，第二阶段为细分割，编码器采用EfficientNet-B5作为主干网，图像多尺度输入；提出了适用于小目标、小样本场景的新损失函数；实验结果表明，该文提出的级联结构网络在颈部淋巴结超声图像分割中网络性能优异，Dice系数达到0.95，较其他UNet方法具有更优的分割性能。     

Segmentation ability map: Interpret deep features for medical image segmentation

Deep convolutional neural networks (CNNs) have been widely used for medical image segmentation. In most studies, only the output layer is exploited to compute the final segmentation results and the hidden representations of the deep learned features have not been well understood. In this paper, we propose a prototype segmentation (ProtoSeg) method to compute a binary segmentation map based on deep features. We measure the segmentation abilities of the features by computing the Dice between the feature segmentation map and ground-truth, named as the segmentation ability score (SA score for short). The corresponding SA score can quantify the segmentation abilities of deep features in different layers and units to understand the deep neural networks for segmentation. In addition, our method can provide a mean SA score which can give a performance estimation of the output on the test images without ground-truth. Finally, we use the proposed ProtoSeg method to compute the segmentation map directly on input images to further understand the segmentation ability of each input image. Results are presented on segmenting tumors in brain MRI, lesions in skin images, COVID-related abnormality in CT images, prostate segmentation in abdominal MRI, and pancreatic mass segmentation in CT images. Our method can provide new insights for interpreting and explainable AI systems for medical image segmentation. Our code is available on: https://github.com/shengfly/ProtoSeg.     

Deep virtual adversarial self-training with consistency regularization for semi-supervised medical image classification

Convolutional neural networks have achieved prominent success on a variety of medical imaging tasks when a large amount of labeled training data is available. However, the acquisition of expert annotations for medical data is usually expensive and time-consuming, which poses a great challenge for supervised learning approaches. In this work, we proposed a novel semi-supervised deep learning method, i.e., deep virtual adversarial self-training with consistency regularization, for large-scale medical image classification. To effectively exploit useful information from unlabeled data, we leverage self-training and consistency regularization to harness the underlying knowledge, which helps improve the discrimination capability of training models. More concretely, the model first uses its prediction for pseudo-labeling on the weakly-augmented input image. A pseudo-label is kept only if the corresponding class probability is of high confidence. Then the model prediction is encouraged to be consistent with the strongly-augmented version of the same input image. To improve the robustness of the network against virtual adversarial perturbed input, we incorporate virtual adversarial training (VAT) on both labeled and unlabeled data into the course of training. Hence, the network is trained by minimizing a combination of three types of losses, including a standard supervised loss on labeled data, a consistency regularization loss on unlabeled data, and a VAT loss on both labeled and labeled data. We extensively evaluate the proposed semi-supervised deep learning methods on two challenging medical image classification tasks: breast cancer screening from ultrasound images and multi-class ophthalmic disease classification from optical coherence tomography B-scan images. Experimental results demonstrate that the proposed method outperforms both supervised baseline and other state-of-the-art methods by a large margin on all tasks.     

Retinal layer segmentation in optical coherence tomography (OCT) using a 3D deep-convolutional regression network for patients with age-related macular degeneration

Introduction – Retinal layer segmentation in optical coherence tomography (OCT) images is an important approach for detecting and prognosing disease. Automating segmentation using robust machine learning techniques lead to computationally efficient solutions and significantly reduces the cost of labor-intensive labeling, which is traditionally performed by trained graders at a reading center, sometimes aided by semi-automated algorithms. Although several algorithms have been proposed since the revival of deep learning, eyes with severe pathological conditions continue to challenge fully automated segmentation approaches. There remains an opportunity to leverage the underlying spatial correlations between the retinal surfaces in the segmentation approach. Methods - Some of these proposed traditional methods can be expanded to utilize the three-dimensional spatial context governing the retinal image volumes by replacing the use of 2D filters with 3D filters. Towards this purpose, we propose a spatial-context, continuity and anatomical relationship preserving semantic segmentation algorithm, which utilizes the 3D spatial context from the image volumes with the use of 3D filters. We propose a 3D deep neural network capable of learning the surface positions of the layers in the retinal volumes. Results - We utilize a dataset of OCT images from patients with Age-related Macular Degeneration (AMD) to assess performance of our model and provide both qualitative (including segmentation maps and thickness maps) and quantitative (including error metric comparisons and volumetric comparisons) results, which demonstrate that our proposed method performs favorably even for eyes with pathological changes caused by severe retinal diseases. The Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) for patients with a wide range of AMD severity scores (0–11) were within 0.84±0.41 and 1.33±0.73 pixels, respectively, which are significantly better than some of the other state-of-the-art algorithms. Conclusion – The results demonstrate the utility of extracting features from the entire OCT volume by treating the volume as a correlated entity and show the benefit of utilizing 3D autoencoder based regression networks for smoothing the approximated retinal layers by inducing shape based regularization constraints.     

Attention-Enriched Deep Learning Model for Breast Tumor Segmentation in Ultrasound Images

Incorporating human domain knowledge for breast tumor diagnosis is challenging because shape, boundary, curvature, intensity or other common medical priors vary significantly across patients and cannot be employed. This work proposes a new approach to integrating visual saliency into a deep learning model for breast tumor segmentation in ultrasound images. Visual saliency refers to image maps containing regions that are more likely to attract radiologists’ visual attention. The proposed approach introduces attention blocks into a U-Net architecture and learns feature representations that prioritize spatial regions with high saliency levels. The validation results indicate increased accuracy for tumor segmentation relative to models without salient attention layers. The approach achieved a Dice similarity coefficient (DSC) of 90.5% on a data set of 510 images. The salient attention model has the potential to enhance accuracy and robustness in processing medical images of other organs, by providing a means to incorporate task-specific knowledge into deep learning architectures.     

Sketch guided and progressive growing GAN for realistic and editable ultrasound image synthesis

Ultrasound (US) imaging is widely used for anatomical structure inspection in clinical diagnosis. The training of new sonographers and deep learning based algorithms for US image analysis usually requires a large amount of data. However, obtaining and labeling large-scale US imaging data are not easy tasks, especially for diseases with low incidence. Realistic US image synthesis can alleviate this problem to a great extent. In this paper, we propose a generative adversarial network (GAN) based image synthesis framework. Our main contributions include: (1) we present the first work that can synthesize realistic B-mode US images with high-resolution and customized texture editing features; (2) to enhance structural details of generated images, we propose to introduce auxiliary sketch guidance into a conditional GAN. We superpose the edge sketch onto the object mask and use the composite mask as the network input; (3) to generate high-resolution US images, we adopt a progressive training strategy to gradually generate high-resolution images from low-resolution images. In addition, a feature loss is proposed to minimize the difference of high-level features between the generated and real images, which further improves the quality of generated images; (4) the proposed US image synthesis method is quite universal and can also be generalized to the US images of other anatomical structures besides the three ones tested in our study (lung, hip joint, and ovary); (5) extensive experiments on three large US image datasets are conducted to validate our method. Ablation studies, customized texture editing, user studies, and segmentation tests demonstrate promising results of our method in synthesizing realistic US images.     

Deep learning provides high accuracy in automated chondrocyte viability assessment in articular cartilage using nonlinear optical microscopy

Chondrocyte viability is a crucial factor in evaluating cartilage health. Most cell viability assays rely on dyes and are not applicable for in vivo or longitudinal studies. We previously demonstrated that two-photon excited autofluorescence and second harmonic generation microscopy provided high-resolution images of cells and collagen structure; those images allowed us to distinguish live from dead chondrocytes by visual assessment or by the normalized autofluorescence ratio. However, both methods require human involvement and have low throughputs. Methods for automated cell-based image processing can improve throughput. Conventional image processing algorithms do not perform well on autofluorescence images acquired by nonlinear microscopes due to low image contrast. In this study, we compared conventional, machine learning, and deep learning methods in chondrocyte segmentation and classification. We demonstrated that deep learning significantly improved the outcome of the chondrocyte segmentation and classification. With appropriate training, the deep learning method can achieve 90% accuracy in chondrocyte viability measurement. The significance of this work is that automated imaging analysis is possible and should not become a major hurdle for the use of nonlinear optical imaging methods in biological or clinical studies.     

Vessel-CAPTCHA: An efficient learning framework for vessel annotation and segmentation

Deep learning techniques for 3D brain vessel image segmentation have not been as successful as in the segmentation of other organs and tissues. This can be explained by two factors. First, deep learning techniques tend to show poor performances at the segmentation of relatively small objects compared to the size of the full image. Second, due to the complexity of vascular trees and the small size of vessels, it is challenging to obtain the amount of annotated training data typically needed by deep learning methods. To address these problems, we propose a novel annotation-efficient deep learning vessel segmentation framework. The framework avoids pixel-wise annotations, only requiring weak patch-level labels to discriminate between vessel and non-vessel 2D patches in the training set, in a setup similar to the CAPTCHAs used to differentiate humans from bots in web applications. The user-provided weak annotations are used for two tasks: (1) to synthesize pixel-wise pseudo-labels for vessels and background in each patch, which are used to train a segmentation network, and (2) to train a classifier network. The classifier network allows to generate additional weak patch labels, further reducing the annotation burden, and it acts as a second opinion for poor quality images. We use this framework for the segmentation of the cerebrovascular tree in Time-of-Flight angiography (TOF) and Susceptibility-Weighted Images (SWI). The results show that the framework achieves state-of-the-art accuracy, while reducing the annotation time by $\sim$77% w.r.t. learning-based segmentation methods using pixel-wise labels for training.     

Semantic instance segmentation with discriminative deep supervision for medical images

Semantic instance segmentation is crucial for many medical image analysis applications, including computational pathology and automated radiation therapy. Existing methods for this task can be roughly classified into two categories: (1) proposal-based methods and (2) proposal-free methods. However, in medical images, the irregular shape-variations and crowding instances (e.g., nuclei and cells) make it hard for the proposal-based methods to achieve robust instance localization. On the other hand, ambiguous boundaries caused by the low-contrast nature of medical images (e.g., CT images) challenge the accuracy of the proposal-free methods. To tackle these issues, we propose a proposal-free segmentation network with discriminative deep supervision (DDS), which at the same time allows us to gain the power of the proposal-based method. The DDS module is interleaved with a carefully designed proposal-free segmentation backbone in our network. Consequently, the features learned by the backbone network become more sensitive to instance localization. Also, with the proposed DDS module, robust pixel-wise instance-level cues (especially structural information) are introduced for semantic segmentation. Extensive experiments on three datasets, i.e., a nuclei dataset, a pelvic CT image dataset, and a synthetic dataset, demonstrate the superior performance of the proposed algorithm compared to the previous works.     

LMISA: A lightweight multi-modality image segmentation network via domain adaptation using gradient magnitude and shape constraint

In medical image segmentation, supervised machine learning models trained using one image modality (e.g. computed tomography (CT)) are often prone to failure when applied to another image modality (e.g. magnetic resonance imaging (MRI)) even for the same organ. This is due to the significant intensity variations of different image modalities. In this paper, we propose a novel end-to-end deep neural network to achieve multi-modality image segmentation, where image labels of only one modality (source domain) are available for model training and the image labels for the other modality (target domain) are not available. In our method, a multi-resolution locally normalized gradient magnitude approach is firstly applied to images of both domains for minimizing the intensity discrepancy. Subsequently, a dual task encoder-decoder network including image segmentation and reconstruction is utilized to effectively adapt a segmentation network to the unlabeled target domain. Additionally, a shape constraint is imposed by leveraging adversarial learning. Finally, images from the target domain are segmented, as the network learns a consistent latent feature representation with shape awareness from both domains. We implement both 2D and 3D versions of our method, in which we evaluate CT and MRI images for kidney and cardiac tissue segmentation. For kidney, a public CT dataset (KiTS19, MICCAI 2019) and a local MRI dataset were utilized. The cardiac dataset was from the Multi-Modality Whole Heart Segmentation (MMWHS) challenge 2017. Experimental results reveal that our proposed method achieves significantly higher performance with a much lower model complexity in comparison with other state-of-the-art methods. More importantly, our method is also capable of producing superior segmentation results than other methods for images of an unseen target domain without model retraining. The code is available at GitHub (https://github.com/MinaJf/LMISA) to encourage method comparison and further research.     

Cross-modal attention for multi-modal image registration

In the past few years, convolutional neural networks (CNNs) have been proven powerful in extracting image features crucial for medical image registration. However, challenging applications and recent advances in computer vision suggest that CNNs are limited in their ability to understand the spatial correspondence between features, which is at the core of image registration. The issue is further exaggerated when it comes to multi-modal image registration, where the appearances of input images can differ significantly. This paper presents a novel cross-modal attention mechanism for correlating features extracted from the multi-modal input images and mapping such correlation to image registration transformation. To efficiently train the developed network, a contrastive learning-based pre-training method is also proposed to aid the network in extracting high-level features across the input modalities for the following cross-modal attention learning. We validated the proposed method on transrectal ultrasound (TRUS) to magnetic resonance (MR) registration, a clinically important procedure that benefits prostate cancer biopsy. Our experimental results demonstrate that for MR-TRUS registration, a deep neural network embedded with the cross-modal attention block outperforms other advanced CNN-based networks with ten times its size. We also incorporated visualization techniques to improve the interpretability of our network, which helps bring insights into the deep learning based image registration methods. The source code of our work is available at https://github.com/DIAL-RPI/Attention-Reg.     

Adaptive convolutional neural networks for accelerating magnetic resonance imaging via k-space data interpolation

Deep learning in k-space has demonstrated great potential for image reconstruction from undersampled k-space data in fast magnetic resonance imaging (MRI). However, existing deep learning-based image reconstruction methods typically apply weight-sharing convolutional neural networks (CNNs) to k-space data without taking into consideration the k-space data's spatial frequency properties, leading to ineffective learning of the image reconstruction models. Moreover, complementary information of spatially adjacent slices is often ignored in existing deep learning methods. To overcome such limitations, we have developed a deep learning algorithm, referred to as adaptive convolutional neural networks for k-space data interpolation (ACNN-k-Space), which adopts a residual Encoder-Decoder network architecture to interpolate the undersampled k-space data by integrating spatially contiguous slices as multi-channel input, along with k-space data from multiple coils if available. The network is enhanced by self-attention layers to adaptively focus on k-space data at different spatial frequencies and channels. We have evaluated our method on two public datasets and compared it with state-of-the-art existing methods. Ablation studies and experimental results demonstrate that our method effectively reconstructs images from undersampled k-space data and achieves significantly better image reconstruction performance than current state-of-the-art techniques. Source code of the method is available at https://gitlab.com/qgpmztmf/acnn-k-space.     

SatCNN: satellite image dataset classification using agile convolutional neural networks

With the launch of various remote-sensing satellites, more and more high-spatial resolution remote-sensing (HSR-RS) images are becoming available. Scene classification of such a huge volume of HSR-RS images is a big challenge for the efficiency of the feature learning and model training. The deep convolutional neural network (CNN), a typical deep learning model, is an efficient end-to-end deep hierarchical feature learning model that can capture the intrinsic features of input HSR-RS images. However, most published CNN architectures are borrowed from natural scene classification with thousands of training samples, and they are not designed for HSR-RS images. In this paper, we propose an agile CNN architecture, named as SatCNN, for HSR-RS image scene classification. Based on recent improvements to modern CNN architectures, we use more efficient convolutional layers with smaller kernels to build an effective CNN architecture. Experiments on SAT data sets confirmed that SatCNN can quickly and effectively learn robust features to handle the intra-class diversity even with small convolutional kernels, and the deeper convolutional layers allow spontaneous modelling of the relative spatial relationships. With the help of fast graphics processing unit acceleration, SatCNN can be trained within about 40 min, achieving overall accuracies of 99.65% and 99.54%, which is the state-of-the-art for SAT data sets.     

Self-supervised pretraining for transferable quantitative phase image cell segmentation

In this paper, a novel U-Net-based method for robust adherent cell segmentation for quantitative phase microscopy image is designed and optimised. We designed and evaluated four specific post-processing pipelines. To increase the transferability to different cell types, non-deep learning transfer with adjustable parameters is used in the post-processing step. Additionally, we proposed a self-supervised pretraining technique using nonlabelled data, which is trained to reconstruct multiple image distortions and improved the segmentation performance from 0.67 to 0.70 of object-wise intersection over union. Moreover, we publish a new dataset of manually labelled images suitable for this task together with the unlabelled data for self-supervised pretraining.