Approximate Size Targets Are Sufficient for Accurate Semantic Segmentation

Deep learning method becomes increasingly popular in many fields, especially in computer vision area, and it is gradually being applied in the medical image area to solve medical image problems such as lesion segmentation, nucleus detection and disease classification. However, training deep learning segment models requires a lot of accurate pixel-level segmentation medical image labels, which is very hard to obtain. Lacking accurate pixel-level segmentation labels is an important factor restricting the development of deep learning method in medical image processing. In this paper, we propose a perivascular dermatitis classification network based on semi-supervised learning which is trained by image-level supervision. Our network can accurately and effectively segment the pathological changes contributed to the disease classification while classifying perivascular dermatitis. We use U-Net as the pathological change discover module to segment the pathological changes, propose the restricted boundary loss to improve the accuracy of the segmentation area boundary and introduce the pathological changes guided module to guide the pathological changes discover module to generate pathological changes contribute to the classification task. We evaluate our network on the dataset of skin pathological images with image-level classification labels. With the image-level labels of the perivascular dermatitis pathological images and normal skin pathological images, our network is trained to discover the pathological changes and classify the skin pathological images. Experiments show that our network discovers the skin pathological changes by using image-level classification labels. The perivascular dermatitis classification accuracy, AUC and PR of our network is 0.8793, 0.8954 and 0.8041.

Transforming the non-structured medical text data into structured data is the basis of the processing and analysis of medical data. The effect of general-purpose word segmentation tools recognizing terminology is not ideal, which greatly affects the accuracy of the word segmentation, and further influences the result of text structuralization. In view of above problems, this paper puts forward a method of discovering new words based on word embedding. It uses Google open source word vector tool word2vec to train text and map the words into abstracted n-dimensional vector space. We can get the latent semantic relations between words and words in the corpus. And then combining the information entropy and word frequency, we can find new words. Finally, we design information extraction rules to get the key information according to the new words, and organize them into structured data. Experimental results on real medical data show that the accuracy is improved by 10% compared to traditional method, and the time is saved by 18% compared to traditional method.

This paper addresses Weakly Supervised Object Localization (WSOL) with only image-level supervision. We propose a Multi-view Learning Localization Network (ML-LocNet) by incorporating multi-view learning into a two-phase WSOL model. The multi-view learning would benefit localization due to the complementary relationships among the learned features from different views and the consensus property among the mined instances from each view. In the first phase, the representation is augmented by integrating features learned from multiple views, and in the second phase, the model performs multi-view co-training to enhance localization performance of one view with the help of instances mined from other views, which thus effectively avoids early fitting. ML-LocNet can be easily combined with existing WSOL models to further improve the localization accuracy. Its effectiveness has been proved experimentally. Notably, it achieves \(68.6\%\) CorLoc and \(49.7\%\) mAP on PASCAL VOC 2007, surpassing the state-of-the-arts by a large margin.

In recent decades, more types and quantities of medical data have been collected due to advanced technology. A large number of significant and critical information is contained in these medical data. High efficient and automated computational methods are urgently needed to process and analyze all available medical data in order to provide the physicians with recommendations and predictions on diagnostic decisions and treatment planning. Traumatic pelvic injury is a severe yet common injury in the United States, often caused by motor vehicle accidents or fall. Information contained in the pelvic Computed Tomography (CT) images is very important for assessing the severity and prognosis of traumatic pelvic injuries. Each pelvic CT scan includes a large number of slices. Meanwhile, each slice contains a large quantity of data that may not be thoroughly and accurately analyzed via simple visual inspection with the desired accuracy and speed. Hence, a computer-assisted pelvic trauma decision-making system is needed to assist physicians in making accurate diagnostic decisions and determining treatment planning in a short period of time. Pelvic bone segmentation is a vital step in analyzing pelvic CT images and assisting physicians with diagnostic decisions in traumatic pelvic injuries. In this study, a new hierarchical segmentation algorithm is proposed to automatically extract multiple-level bone structures using a combination of anatomical knowledge and computational techniques. First, morphological operations, image enhancement, and edge detection are performed for preliminary bone segmentation. The proposed algorithm then uses a template-based best shape matching method that provides an entirely automated segmentation process. This is followed by the proposed Registered Active Shape Model (RASM) algorithm that extracts pelvic bone tissues using more robust training models than the Standard ASM algorithm. In addition, a novel hierarchical initialization process for RASM is proposed in order to address the shortcoming of the Standard ASM, i.e. high sensitivity to initialization. Two suitable measures are defined to evaluate the segmentation results: Mean Distance and Mis-segmented Area to quantify the segmentation accuracy. Successful segmentation results indicate effectiveness and robustness of the proposed algorithm. Comparison of segmentation performance is also conducted using both the proposed method and the Snake method. A cross-validation process is designed to demonstrate the effectiveness of the training models. 3D pelvic bone models are built after pelvic bone structures are segmented from consecutive 2D CT slices. Automatic and accurate detection of the fractures from segmented bones in traumatic pelvic injuries can help physicians detect the severity of injuries in patients. The extraction of fracture features (such as presence and location of fractures) as well as fracture displacement measurement, are vital for assisting physicians in making faster and more accurate decisions. In this project, after bone segmentation, fracture detection is performed using a hierarchical algorithm based on wavelet transformation, adaptive windowing, boundary tracing and masking. Also, a quantitative measure of fracture severity based on pelvic CT scans is defined and explored. The results are promising, demonstrating that the proposed method not only capable of automatically detecting both major and minor fractures, but also has potentials to be used for clinical applications.

In order to improve the accuracy of image segmentation without changing the structure of original semantic segmentation models, an approach to train semantic segmentation models by using generative adversarial network (SS-GAN) is proposed. Using adversarial network to distinguish the source of segmented images, the model can learn the high-order relationship between pixels to enhance the spatial continuity of pixels in the segmented image. There are three aspects related to the work: constructing the generative model of fully convolutional network (FCN) structure by using VGG, and segment image preliminarily; constructing the adversarial model and training it by combining the original images, fake segmented images and real segmented images; modifying the loss function, adding the anti-loss to assist segmentation model training, encouraging generative network to learn the inter-pixel relationship independently. Experiments on PASCAL VOC and Cityscapes datasets show that the proposed method achieves better performance than the existing advanced methods, and improves IoU by 1.56% and 1.93%, respectively.

A neurodegenerative illness known as Alzheimer's causes the loss of brain cells and the progressive atrophy of brain tissue. It badly affects a person’s normal life. However, if we are able to detect it early and treat it, most patients will be able to recover to some degree and lead a normal life with some dependence. Continuous clinical assessment is needed for diagnosing this type of disorder. Medical diagnosis today extensively relies on deep learning approaches. However, medical image data analysis has a lot of constraints. One of the major constraints faced during medical image analysis is data scarcity and data imbalance. In light of these concerns, the current study sets out to create a hybrid deep learning model that can effectively categorise various Alzheimer's disease variants using magnetic resonance imaging (MRI) data. For solving data imbalance, first, we blur and sharpen all the images, and finally, we pass all these images along with the original images through a predefined CNN (Convolutional Neural Network) model that was trained using mnist weight for extracting features, then pass these features to an extra-tree classifier for feature reduction, and finally input these reduced features to a customised deep learning model. This work used different pre-trained models for extracting features for our customised DNN (Deep Neural Network) and compared those models with the cutting-edge model chosen as the base model. The results state that our proposed model, which was pre-trained using ResNet with a dropout concept, got the highest values of training accuracy (98.20) and validation accuracy (92.61). This model also effectively addresses the problem of overfitting.

Sepsis is a severe life‐threatening disease, which is the body's extreme reverberation to any virus leading to failed organs, damaged tissue, or death. It requires accurate and efficient real‐time detection. Continuous and potent monitoring of patient health data can be useful in predicting the potential risks the patient might be exposed to, based on their recent history of medical records. Machine learning models have proven to be a significant approach in performing accurate and precise predictions, especially in the medical field. In this paper, a smart network with a long short term memory (LSTM) based model at its heart is proposed for early prediction of Sepsis for patients admitted in the ICU. The network operates on time series data and predicts the probability of a patient developing sepsis based on the historical medical data of the patient. It comprises Internet of Things based devices which have proven to be impactful in the areas of acquiring continuous and real‐time data in the healthcare field. In this paper, an architecture for early prediction and monitoring Sepsis with minimized latency through LSTM network is proposed by designing a decentralized prediction model in a fog‐based environment. The model had an accuracy of 95.1% after the last epoch with validation accuracy as 95%. The receiver operating characteristic curve area was reported as 0.864 for testing data with an accuracy of 95.1%. The model was not over‐fitted since the validation accuracy was significantly close to the training accuracy of the model.

Simple low-cost approaches to semantic segmentation in radiation therapy planning for prostate cancer using deep learning with non-contrast planning CT images.

Histological analysis with microscopy is the gold standard to diagnose and stage cancer, where slides or whole slide images are analyzed for cell morphological and spatial features by pathologists. The nuclei of cancerous cells are characterized by nonuniform chromatin distribution, irregular shapes, and varying size. As nucleus area and shape alone carry prognostic value, detection and segmentation of nuclei are among the most important steps in disease grading. However, evaluation of nuclei is a laborious, time-consuming, and subjective process with large variation among pathologists. Recent advances in digital pathology have allowed significant applications in nuclei detection, segmentation, and classification, but automated image analysis is greatly affected by staining factors, scanner variability, and imaging artifacts, requiring robust image preprocessing, normalization, and segmentation methods for clinically satisfactory results. In this paper, we aimed to evaluate and compare the digital image analysis techniques used in clinical pathology and research in the setting of gastric cancer. A literature review was conducted to evaluate potential methods of improving nuclei detection. Digitized images of 35 patients from a retrospective cohort of gastric adenocarcinoma at Oulu University Hospital in 1987–2016 were annotated for nuclei (n = 9085) by expert pathologists and 14 images of different cancer types from public TCGA dataset with annotated nuclei (n = 7000) were used as a comparison to evaluate applicability in other cancer types. The detection and segmentation accuracy with the selected color normalization and stain separation techniques were compared between the methods. The extracted information can be supplemented by patient’s medical data and fed to the existing statistical clinical tools or subjected to subsequent AI-assisted classification and prediction models. The performance of each method is evaluated by several metrics against the annotations done by expert pathologists. The F1-measure of 0.854 ± 0.068 is achieved with color normalization for the gastric cancer dataset, and 0.907 ± 0.044 with color deconvolution for the public dataset, showing comparable results to the earlier state-of-the-art works. The developed techniques serve as a basis for further research on application and interpretability of AI-assisted tools for gastric cancer diagnosis.

The research aims to develop an advanced and precise lung cancer screening model based on Convolutional Neural Networks (CNN). Based on the health medical big data platform of Shandong University, we developed a VGG16-Based CNN lung cancer screening model. This model was trained using the Computed Tomography scans data of patients from Pingyi Traditional Chinese Medicine Hospital in Shandong Province, from January to February 2023. Data augmentation techniques, including random resizing, cropping, horizontal flipping, color jitter, random rotation and normalization, were applied to improve model generalization. We used five-fold cross-validation to robustly assess performance. The model was fine-tuned with an SGD optimizer (learning rate 0.001, momentum 0.9, and L2 regularization) and a learning rate scheduler. Dropout layers were added to prevent the model from relying too heavily on specific neurons, enhancing its ability to generalize. Early stopping was implemented when validation loss did not decrease over 10 epochs. In addition, we evaluated the model's performance with Area Under the Curve (AUC), Classification accuracy, Positive Predictive Value (PPV), and Negative Predictive Value (NPV), Sensitivity, Specificity and F1 score. External validation used an independent dataset from the same hospital, covering January to February 2022. The training and validation loss and accuracy over iterations show that both accuracy metrics peak at over 0.9 by iteration 15, prompting early stopping to prevent overfitting. Based on five-fold cross-validation, the ROC curves for the VGG16-Based CNN model, demonstrate an AUC of 0.963 ± 0.004, highlighting its excellent diagnostic capability. Confusion matrices provide average metrics with a classification accuracy of 0.917 ± 0.004, PPV of 0.868 ± 0.015, NPV of 0.931 ± 0.003, Sensitivity of 0.776 ± 0.01, Specificity of 0.962 ± 0.005 and F1 score of 0.819 ± 0.008, respectively. External validation confirmed the model's robustness across different patient populations and imaging conditions. The VGG16-Based CNN lung screening model constructed in this study can effectively identify lung tumors, demonstrating reliability and effectiveness in real-world medical settings, and providing strong theoretical and empirical support for its use in lung cancer screening.

Specific medical data has limitations in that there are not many numbers and it is not standardized. to solve these limitations, it is necessary to study how to efficiently process these limited amounts of data. In this paper, deep learning methods for automatically determining cardiovascular diseases are described, and an effective preprocessing method for CT images that can be applied to improve the performance of deep learning was conducted. The cardiac CT images include several parts of the body such as the heart, lungs, spine, and ribs. The preprocessing step proposed in this paper divided CT image data into regions of interest and other regions using K-means clustering and the Grabcut algorithm. We compared the deep learning performance results of original data, data using only K-means clustering, and data using both K-means clustering and the Grabcut algorithm. All data used in this paper were collected at Soonchunhyang University Cheonan Hospital in Korea and the experimental test proceeded with IRB approval. The training was conducted using Resnet 50, VGG, and Inception resnet V2 models, and Resnet 50 had the best accuracy in validation and testing. Through the preprocessing process proposed in this paper, the accuracy of deep learning models was significantly improved by at least 10% and up to 40%.

With the development of deep learning, convolutional neural networks have become the mainstream of computer vision algorithms. In recent years, the biggest problem of applying convolutional neural networks to image segmentation is that they cannot achieve accurate segmentation of the last layer, also cause resolution loss when extracting features, and cannot meet the demand of different pixels requiring different context dependencies. To address these issues, we add an attention mechanism and a jump feature fusion method to deeplabv3+ so that features are extracted without severe feature loss and a broader range of contextual information can be encoded into local features. The feature map is further enriched by adding a module combining bilinear upsampling and deconvolution in the process of feature restoration. Compared to previous algorithms, the results of this algorithm are superior. A performance of 85.73% is achieved on PASCAL VOC2012 using the proposed model.

In the systems of industrial robotics and autonomous vehicles, instance segmentation is widely employed. However, manually labelling an object outline is time‐consuming. In order to reduce annotation costs, we present a weakly supervised instance segmentation method in this article. A deeply convolutional network is first used to construct multi‐scale feature maps for each object in the input image. After that, the encoder‐decoder framework with dynamic convolution is utilised to enhance model capacity and efficiency, while avoiding the issues of anchor design, proposal selection, and RoIAlign implementation. In particular, Dynamic Heads are used in the encoder to create dynamic convolution kernels, while Instance Heads are used in the decoder to provide the global feature map. With dynamic convolution, each instance can be segmented independently, reducing interference with other instances and improving segmentation accuracy. Under the supervision of projection loss and pixel point colour pairing loss, the contours of each object are finally outlined. On the PASCAL VOC and MS COCO datasets, the proposed method is competitive with more sophisticated approaches. In the VOC dataset, segmentation performance achieved 37.6% average precision with ResNet‐101 and FPN networks. The extensively visualised results demonstrate the effectiveness of the proposed encoder‐decoder framework with dynamic convolution.

Recent advancements in deep learning have greatly improved computer vision tasks like object detection, image classification, and segmentation. Despite these successes, traditional supervised learning methods still depend on large annotated datasets, which are often expensive and time consuming to create. To overcome this limitation, we present a zero-shot segmentation framework that combines the strengths of CLIP (Contrastive Language-Image Pretraining), its segmentation-focused variant CLIPSeg, and the Segment Anything Model (SAM). This approach first uses the zero-shot classification ability of CLIP or CLIPSeg to produce initial segmentation cues. These cues, such as point and box prompts, are then refined by SAM to generate accurate segmentation masks. Using this prompt-based strategy, the system can perform segmentation without requiring labeled data, making it suitable for a wide range of domains, including both natural scenes and medical imaging. Our experiments on benchmarks such as MS-COCO, Pascal VOC, and chest X-ray datasets highlight the effectiveness of the method. In particular, the CLIPSeg+SAM combination achieves a mean IoU of 0.793 and a Dice score of 0.873 in the chest X-ray dataset, outperforming both CLIPSeg and SAM when used alone. Visual results also show that this method produces clearer and more precise mask boundaries, even in challenging or cluttered environments. In summary, the proposed training-free framework offers a scalable and generalizable solution for zero-shot segmentation, significantly reducing the reliance on annotated datasets while delivering strong performance on unseen classes.

Real-time semantic segmentation is widely used in autonomous driving and robotics. Most previous networks achieved great accuracy based on a complicated model involving mass computing. The existing lightweight networks generally reduce the parameter sizes by sacrificing the segmentation accuracy. It is critical to balance the parameters and accuracy for real-time semantic segmentation. In this article, we propose a lightweight multiscale-feature-fusion network (LMFFNet) mainly composed of three types of components: split-extract-merge bottleneck (SEM-B) block, feature fusion module (FFM), and multiscale attention decoder (MAD), where the SEM-B block extracts sufficient features with fewer parameters. FFMs fuse multiscale semantic features to effectively improve the segmentation accuracy and the MAD well recovers the details of the input images through the attention mechanism. Without pretraining, LMFFNet-3-8 achieves 75.1% mean intersection over union (mIoU) with 1.4 M parameters at 118.9 frames/s using RTX 3090 GPU. More experiments are investigated extensively on various resolutions on other three datasets of CamVid, KITTI, and WildDash2. The experiments verify that the proposed LMFFNet model makes a decent tradeoff between segmentation accuracy and inference speed for real-time tasks. The source code is publicly available at https://github.com/Greak-1124/LMFFNet.