Self-supervised learning based transformer and convolution hybrid network for one-shot organ segmentation

One-shot organ segmentation (OS2) aims at segmenting the desired organ regions from the input medi-cal imaging data with only one pre-annotated example as the reference. By using the minimal annotation data to facilitate organ segmentation, OS2 receives great attention in the medical image analysis community due to its weak re-quirement on human annotation. In OS2, one core issue is to explore the mutual information between the support (ref-erence slice) and the query (test slice). Existing methods rely heavily on the similarity between slices, and additional slice allocation mechanisms need to be designed to reduce the impact of the similarity between slices on the segmen-tation performance. To address this issue, we build a novel support-query interactive embedding (SQIE) module, which is equipped with the channel-wise co-attention, spatial-wise co-attention, and spatial bias transformation blocks to identify "what to look", "where to look", and "how to look" in the input test slice. By combining the three mechanisms, we can mine the interactive information of the intersection area and the disputed area between slices, and establish the feature connection between the target in slices with low similarity. We also propose a self-supervised contrastive learning framework, which transforms knowledge from the physical position to the embedding space to facilitate the self-supervised interactive embedding of the query and support slices. Comprehensive experiments on two large benchmarks demonstrate the superior capacity of the pro-posed approach when compared with the current alterna-tives and baseline models.

In this study, we investigated how shapes are classified based on local and global features by four representative convolutional neural networks (CNNs), i.e., AlexNet, VGG, ResNet and Inception. While the local features are based on simple components, such as orientation of line segment, the global features are based on the whole object, such as whether an object has a hole. For example, solid triangles and solid squares are differentiated by local features, solid circles and rings are differentiated by a global feature. Two sets of experiments were performed in this research. In the first experiment, we examined how the four CNNs pre-trained on ImageNet (with transfer learning) learned to differentiate the regular shapes (equilateral-triangles, squares, circles and rings). Our results showed that the pre-trained CNNs exhibited faster learning rates in the tasks discriminating the local features than in the tasks discriminating the global feature. However, the transfer learning of discriminating the global feature in regular shapes were better generalized to irregular shapes than the transfer learning of discriminating local features. In the second experiment, the CNNs were trained from scratch (with random weights initialization) to discriminate local and global features in regular and irregular shapes. Different from the transfer learning, the CNNs exhibited faster learning rates in discriminating the global feature than the local features. Similar to transfer learning, the CNNs exhibited excellent learning generalization to discriminating the global feature of irregular shapes, but poor learning generalization to discriminating the local features in the irregular shapes. The overarching goal of this research is to create a paradigm and benchmark to directly compare how the CNNs and primate visual systems process geometrical invariants. In contrast to the ImageNet approach which employs natural images to train CNNs, we employed the “ShapeNet” approach which features geometrical shapes with well-defined properties. The ShapeNet approach will not only help elucidate the strengths and limitations of CNN computation, but also provide insights into visual information processing in the primates.

Traditional multimodal contrastive learning brings text and its corresponding image closer together as a positive pair, where the text typically consists of fixed sentence structures or specific descriptive statements, and the image features are generally global features (with some fine-grained work using local features). Similar to unimodal self-supervised contrastive learning, this approach can be seen as enforcing a strict identity constraint in a multimodal context. However, due to the inherent complexity of remote sensing images, which cannot be easily described in a single sentence, and the fact that remote sensing images contain rich ancillary information beyond just object features, this strict identity constraint may be insufficient. To fully leverage the characteristics of remote sensing images, we propose a multimodal contrastive learning method for remote sensing image feature extraction, based on positive sample tripartite relaxation, where the model is relaxed in three aspects. The first aspect of relaxation involves both the text and image inputs. By introducing learnable parameters in the language and image branches, instead of relying on fixed sentence structures and fixed image features, the network can achieve a more flexible description of remote sensing images in text and extract ancillary information from the image features, thereby relaxing the input constraints. Second relaxation is achieved through multimodal alignment of various features. By aligning semantic information with the corresponding semantic regions in the images, the method allows for the relaxation of local image features under semantic constraints. This approach addresses the issue of selecting image patches in unimodal settings, where there is no semantic constraint. The proposed method for remote sensing image feature extraction has been validated on four datasets. On the PatternNet dataset, it achieved a 91.1% accuracy with just one-shot.

Spatiotemporal self-supervised predictive learning for atmospheric variable prediction via multi-group multi-attention

[18F]-Fluorodeoxyglucose (FDG) positron emission tomography - computed tomography (PET-CT) has become the imaging modality of choice for diagnosing many cancers. Co-learning complementary PET-CT imaging features is a fundamental requirement for automatic tumor segmentation and for developing computer aided cancer diagnosis systems. In this study, we propose a hyper-connected transformer (HCT) network that integrates a transformer network (TN) with a hyper connected fusion for multi-modality PET-CT images. The TN was leveraged for its ability to provide global dependencies in image feature learning, which was achieved by using image patch embeddings with a self-attention mechanism to capture image-wide contextual information. We extended the single-modality definition of TN with multiple TN based branches to separately extract image features. We also introduced a hyper connected fusion to fuse the contextual and complementary image features across multiple transformers in an iterative manner. Our results with two clinical datasets show that HCT achieved better performance in segmentation accuracy when compared to the existing methods.Clinical Relevance-We anticipate that our approach can be an effective and supportive tool to aid physicians in tumor quantification and in identifying image biomarkers for cancer treatment.

Local feature extraction is one of the key characteristics of convolutional neural networks (CNNs). This study proposes an adaptive local feature enhancement (ALFE) model with a low‐frequency general appearance‐enhancement operator and a high‐frequency local detail enhancement operator to improve local features of CNNs. Through supervised training, the model could adaptively adjust enhancement parameters and achieve a global‐local enhancement of training images and CNNs. The performance of ALFE was first preliminarily evaluated with a self‐built CNN on CIFAR‐10 data set in different conditions of image augmentation and feature pooling. CNNs with ALFE could increase the top‐1 accuracy compared against CNNs in the same conditions, and nearly reach the same level of performance for different pooling approaches. With only two extra adjustable parameters, this model could effectively avoid overfitting, without affecting the convergence speed of CNN. In addition, the extra burden of network complexity could be neglected. Further, experiments of three existing popular CNNs (AlexNet, VGGNet and ResNet) with ALFE were carried out on dogs versus cats, Tiny ImageNet, and SVHN data sets, respectively. The results show that ALFE is feasible for the existing popular CNN models, improving their top‐1 accuracies without changing their convergence speeds.

Whether tracking patient progress for clinical decision making, or investigating novel therapies, automated analysis of Computed Tomography (CT) imaging data is essential for the future of digital radiomics. In digital radiomics, as in all medical imaging, well annotated data is scarce, whereas unlabelled images are relatively plentiful. In other fields of image processing and medical imaging, the application of self-supervised learning (SSL) to large quantities of unlabelled data has resulted in great strides forward for fast, scalable, and interpretable image analysis. In this work we present a new approach applying SSL to CT imaging data which allows for: 1) Improved performance on image classification tasks, based on 2) dramatically reduced quantity of annotated CT imaging data, whilst also 3) enabling easy exploration and interpretation of the image regions. We applied a selection of self-supervised approaches (BYOL, DINO, SimCLR, & inpainting) to CT imaging data. Because the high dimensionality of CT data prevents us from using them directly to out-of the box SSL models, we adopt a 3D patching approach to reduce the dimensionality of the neural net input, and process each patch independently. We train our self-supervised models on public datasets (DeepLesion, NSCLC), and we specialize these models for tumor classification tasks that we evaluate on AstraZeneca sponsored clinical trials. The specialization is done in two different ways: 1) we use the pre-trained SSL model as an encoder that transforms the images of the clinical study into embeddings, on which we apply supervised classification models; 2) we use transfer learning to fine-tune supervised classification models that take the patches directly as inputs. We find that self-supervised pre-training significantly improves the accuracy on tumor classification tasks compared against a supervised learning baseline. Additionally, using the SSL embeddings we build an interactive map of CT imaging data enabling quick and intuitive inspection of the relevant regions. Our findings show that SSL constitutes an important tool for medical imaging analysis. SSL results in models that generalize better, and enable improved downstream interpretability and predictions. Furthermore, well trained SSL models can be re-applied to multiple indications because they are pre-trained on broad and diverse CT imaging data. Citation Format: Leon Fedden, Zhenning Zhang, Khan Baykaner, Qin Li, Lucas Bordeaux. DIME-CT: Self-supervised learning for medical image analysis using patch-based embeddings [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1937.

Recent decades have witnessed rapid development in the field of medical image segmentation. Deep learning-based fully convolution neural networks have played a significant role in the development of automated medical image segmentation models. Though immensely effective, such networks only take into account localized features and are unable to capitalize on the global context of medical image. In this paper, two deep learning based models have been proposed namely USegTransformer-P and USegTransformer-S. The proposed models capitalize upon local features and global features by amalgamating the transformer-based encoders and convolution-based encoders to segment medical images with high precision. Both the proposed models deliver promising results, performing better than the previous state of the art models in various segmentation tasks such as Brain tumor, Lung nodules, Skin lesion and Nuclei segmentation. The authors believe that the ability of USegTransformer-P and USegTransformer-S to perform segmentation with high precision could remarkably benefit medical practitioners and radiologists around the world.

We investigated the application of multiresolution global and local texture features to reduce false-positive detection in a computerized mass detection program. One hundred and sixty-eight digitized mammograms were randomly and equally divided into training and test groups. From these mammograms, two datasets were formed. The first dataset (manual) contained four regions of interest (ROIs) selected manually from each of the mammograms. One of the four ROIs contained a biopsy-proven mass and the other three contained normal parenchyma, including dense, mixed dense/fatty, and fatty tissues. The second dataset (hybrid) contained the manually extracted mass ROIs, along with normal tissue ROIs extracted by an automated Density-Weighted Contrast Enhancement (DWCE) algorithm as false-positive detections. A wavelet transform was used to decompose an ROI into several scales. Global texture features were derived from the low-pass coefficients in the wavelet transformed images. Local texture features were calculated from the suspicious object and the peripheral subregions. Linear discriminant models using effective features selected from the global, local, or combined feature spaces were established to maximize the separation between masses and normal tissue. Receiver Operating Characteristic (ROC) analysis was conducted to evaluate the classifier performance. The classification accuracy using global features were comparable to that using local features. With both global and local features, the average area, Az, under the test ROC curve, reached 0.92 for the manual dataset and 0.96 for the hybrid dataset, demonstrating statistically significant improvement over those obtained with global or local features alone. The results indicated the effectiveness of the combined global and local features in the classification of masses and normal tissue for false-positive reduction.

In order to overcome the effects of posture, illumination, expression and other factors on face recognition, this paper proposes an algorithm which is based on local binary pattern (LBP) and convolutional neural network (CNN). LBP is a texture description method which describes the local texture features of an image. It has good robustness of illumination and posture. CNN can effectively extract the spatial features of images and reduce the dimensions of features. This paper combines the advantages of LBP and CNN to improve the accuracy of face recognition. The CNN in this algorithm has four convolution layers, two max-pooling layers, one activation layer, one fully connected layer, and one output layer. In order to optimize the network structure, batch normalization layer is added after the convolution layer. We get the local binary pattern coded images and put the images as the input of the CNN and train the network. Hence, we can use the well-trained CNN for classification and identification. The Experiments on the CMU-PIE face database show that our algorithm can effectively improve the rate of face recognition.

DFENet: A dual-branch feature enhanced network integrating transformers and convolutional feature learning for multimodal medical image fusion

Surveillance cameras can be deployed in various environments where lighting conditions are constantly changing. However, due to the limited dynamic range of current image sensors, the captured images are only low dynamic range images that usually suffer from over-exposure and under-exposure situations where important details are lost. Therefore, it is critical to recover the lost details of such images in order to improve visual experience for observers and performance for possible computer vision processing. In this paper, we propose a reformulated Laplacian pyramid and a convolutional neural network (CNN) model to enhance and recover the lost detail of a degraded image. The reformulated Laplacian first decomposes the image into two sub-images that contain global and local image features, respectively. The global features and local features are processed by the proposed CNN model to manipulate the global luminance terrain and enhance local details. The final image is obtained by reconstructing the CNN generated local and global features. Various experiments have been conducted. The results demonstrate that the proposed model outperforms the state-of-the-art methods.

In recent years, significant progress has been made in crowd counting with the development of convolutional neural networks (CNNs). However, while CNNs excel at extracting local features, their limited receptive fields restrict their ability to model global context. In contrast, Transformers can effectively model long-distance dependencies, but are inferior to CNN in capturing local detail features. Local details and global context information are crucial to handle large-scale changes in crowds. To address this problem, we propose a novel dual backbone network (DBNet) that integrates CNN and Transformer architectures, aiming to capture and aggregate both global semantic information and local detail features at multiple levels. Specifically, the dual backbone structure is designed to extract fine-grained local features while modeling long-range contextual relationships. Additionally, we introduce a multi-attention hierarchical fusion module (MAHF) that integrates global and local features from the two backbones while suppressing background noise. To further enhance accuracy in the presence of multi-scale variations, we also employ a Feature Enhancement Module (FEM), which enables the network to more effectively identify edge features and facilitates more effective multi-scale feature modeling. Extensive experiments on ShanghaiTech, UCF-QNRF, and JHU-Crowd++ datasets demonstrate that DBNet achieves competitive performance, validating the effectiveness of our approach.

Lung cancer constitutes the most severe cause of cancer-related mortality. Recent evidence supports that early detection by means of computed tomography (CT) scans significantly reduces mortality rates. Given the remarkable progress of Vision Transformers (ViTs) in the field of computer vision, we have delved into comparing the performance of ViTs versus Convolutional Neural Networks (CNNs) for the automatic identification of lung cancer based on a dataset of 212 medical images. Importantly, neither ViTs nor CNNs require lung nodule annotations to predict the occurrence of cancer. To address the dataset limitations, we have trained both ViTs and CNNs with three advanced techniques: transfer learning, self-supervised learning, and sharpness-aware minimizer. Remarkably, we have found that CNNs achieve highly accurate prediction of a patient’s cancer status, with an outstanding recall (93.4%) and area under the Receiver Operating Characteristic curve (AUC) of 98.1%, when trained with self-supervised learning. Our study demonstrates that both CNNs and ViTs exhibit substantial potential with the three strategies. However, CNNs are more effective than ViTs with the insufficient quantities of dataset.

paper, we focus on exploring the relationship between weave patterns and their mechanical properties in woven fiber composites through Machine Learning. Specifically, we explore the interactions between woven architectures and in-plane stiffness properties through Deep Convolutional Neural Network (DCNN) and Generative Adversarial Network (GAN). Our research is important for exploring how woven composite’s pattern is related to its mechanical properties and accelerating woven composite design as well as optimization. We focus on two tasks: (1) Stiffness prediction: Predicting in-plane stiffness properties for given weave patterns. Our DCNN extracts high-level features through several convolutional and fully connected layers to determine the final predictions. (2) Weave pattern prediction: Predicting weave patterns for target stiffness properties, which can be treated as the reverse task of the first one. Due to many-to-one mapping between weave patterns and the composite properties, we utilize a Decoder Neural Network as our baseline model and compare its performance with GAN and Genetic Algorithm. We represent the weave patterns as 2D checkerboard models and use finite element analysis (FEA) to determine in-plane stiffness properties, which serve as input data for our ML framework. We show that: (1) for stiffness prediction, DCNN can predict stiffness values for a given weave pattern with relatively high accuracy (above 93%); (2) for weave pattern prediction, the GAN model gives the best prediction accuracy (above 92%) while Decoder Neural Network has the best time efficiency. HAOTIAN FENG