Segmentation Results Research Articles

Cattle Body Size Measurement Based on DUOS-PointNet+.

The common non-contact, automatic body size measurement methods based on the whole livestock point cloud are complex and prone to errors. Therefore, a cattle body measuring system is proposed. The system includes a new algorithm called dynamic unbalanced octree grouping (DUOS), based on PointNet++, and an efficient method of body size measurement based on segmentation results. This system is suitable for livestock body feature sampling. The network divides the cow into seven parts, including the body and legs. Moreover, the key points of body size are located in the different parts. It combines density measurement, point cloud slicing, contour extraction, point cloud repair, etc. A total of 137 items of cattle data are collected. Compared with some of the other models, the DUOS algorithm improves the accuracy of the segmentation task and mean intersection by 0.53% and 1.21%, respectively. Moreover, compared with the manual measurement results, the relative errors of the experimental measurement results are as follows: withers height, 1.18%; hip height, 1.34%; body length, 2.52%; thoracic circumference, 2.12%; abdominal circumference, 2.26%; and cannon circumference, 2.78%. In summary, the model is proven to have a good segmentation effect on cattle bodies and is suitable for cattle body size measurement.

Automatic segmentation of esophageal cancer, metastatic lymph nodes and their adjacent structures in CTA images based on the UperNet Swin network.

To create a deep-learning automatic segmentation model for esophageal cancer (EC), metastatic lymph nodes (MLNs) and their adjacent structures using the UperNet Swin network and computed tomography angiography (CTA) images and to improve the effectiveness and precision of EC automatic segmentation and TN stage diagnosis. Attention U-Net, UperNet Swin, UNet++ and UNet were used to train the EC segmentation model to automatically segment the EC, esophagus, pericardium, aorta and MLN from CTA images of 182 patients with postoperative pathologically proven EC. The Dice similarity coefficient (DSC), sensitivity, and positive predictive value (PPV) were used to assess their segmentation effectiveness. The volume of EC was calculated using the segmentation results, and the outcomes and times of automatic and human segmentation were compared. All statistical analyses were completed using SPSS 25.0 software. Among the four EC autosegmentation models, the UperNet Swin had the best autosegmentation results with a DSC of 0.7820 and the highest values of EC sensitivity and PPV. The esophagus, pericardium, aorta and MLN had DSCs of 0.7298, 0.9664, 0.9496 and 0.5091. The DSCs of the UperNet Swin were 0.6164, 0.7842, 0.8190, and 0.7259 for T1-4 EC. The volume of EC and its adjacent structures between the ground truth and UperNet Swin model were not significantly different. The UperNet Swin showed excellent efficiency in autosegmentation and volume measurement of EC, MLN and its adjacent structures in different T stage, which can help to T and N stage diagnose EC and will save clinicians time and energy.

Physics-guided interpretable CNN for SAR target recognition

Deep Learning (DL) model has been widely used in the field of Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR) and has achieved excellent performance. However, the black-box nature of DL models has been the focus of criticism, especially in the application of SAR-ATR, which is closely associated with the national defense and security domain. To address these issues, a new interpretable recognition model Physics-Guided BagNet (PGBN) is proposed in this article. The model adopts an interpretable convolutional neural network framework and uses time–frequency analysis to extract physical scattering features in SAR images. Based on the physical scattering features, an unsupervised segmentation method is proposed to distinguish targets from the background in SAR images. On the basis of the segmentation result, a structure is designed, which constrains the model’s spatial attention to focus more on the targets themselves rather than the background, thereby making the model’s decision-making more in line with physical principles. In contrast to previous interpretable research methods, this model combines interpretable structure with physical interpretability, further reducing the model’s risk of error recognition. Experiments on the MSTAR dataset verify that the PGBN model exhibits excellent interpretability and recognition performance, and comparative experiments with heatmaps indicate that the physical feature guidance module presented in this article can constrain the model to focus more on the target itself rather than the background.

Hybrid clinical-radiomics model based on fully automatic segmentation for predicting the early expansion of spontaneous intracerebral hemorrhage: A multi-center study

BackgroundEarly prediction of hematoma expansion (HE) is important for the development of therapeutic strategies for spontaneous intracerebral hemorrhage (sICH). Radiomics can help to predict early hematoma expansion in intracerebral hemorrhage. However, complex image processing procedures, especially hematoma segmentation, are time-consuming and dependent on assessor experience. We provide a fully automated hematoma segmentation method, and construct a hybrid predictive model for risk stratification of hematoma expansion. PurposeTo propose an automatic approach for predicting early hemorrhage expansion after spontaneous intracerebral hemorrhage using deep-learning and radiomics methods. MethodsA total of 258 patients with sICH were retrospectively enrolled for model construction and internal validation, while another two cohorts (n=87, 149) were employed for independent validation. For hemorrhage segmentation, an iterative segmentation procedure was performed to delineate the area using an nnU-Net framework. Radiomics models of intra-hemorrhage and multiscale peri-hemorrhage were established and evaluated, and the best discriminative-scale peri-hemorrhage radiomics model was selected for further analysis. Combining clinical factors and intra- and peri-hemorrhage radiomics signatures, a hybrid nomogram was constructed for the early HE prediction using multivariate logistic regression. For model validation, the receiver operating characteristic (ROC) curve analyses and DeLong test were used to evaluate the performances of the constructed models, and the calibration curve and decision curve analysis were performed for clinical application. ResultsOur iterative auto-segmentation model showed satisfactory results for hematoma segmentation in all four cohorts. The Dice similarity coefficient of this hematoma segmentation model reached 0.90, showing an expert-level accuracy in hematoma segmentation. The consumed time of the efficient delineation was significantly decreased, from 18 min to less than 2 min, with the assistance of the auto-segmentation model. The radiomics model of 2-mm peri-hemorrhage had a preferable area under ROC curve (AUC) of 0.840 (95 % confidence interval [CI]: 0.768, 0.912) compared with the original (0-mm dilatation) model with an AUC of 0.796 (95 % CI: 0.717, 0.875). The clinical–radiomics hybrid model showed better performances for HE prediction, with AUC of 0.853, 0.852, 0.772, and 0.818 in the training, internal validation, and independent validation cohorts 1 and 2, respectively. ConclusionsThe fully automatic clinical–radiomics model based on deep learning and radiomics exhibits a good ability for hematoma segmentation and a favorable performance in stratifying HE risks.

FaultSeg3D plus: A comprehensive study on evaluating and improving CNN-based seismic fault segmentation

Convolutional neural networks (CNNs) have been widely used for seismic fault segmentation and show more powerful performance than conventional attribute-based methods to obtain a fault map with noise-free and continuously trackable fault features. However, CNN-based methods face the potential problem of poor generalization in field seismic images, and factors affecting fault segmentation remain incompletely studied or unexplored. Moreover, the existing pixel-wise metrics, borrowed from the natural image segmentation tasks, cannot fairly or reasonably evaluate the fault segmentation results. We first develop a distance-based metric to provide a geologically more reasonable evaluation on fault interpretation. We then use the most commonly used U-net architecture as an example to study how the CNN-based fault segmentation is affected by some significant factors such as training data, many kinds of network hyperparameters, and scaling and rotation in the inference step. Experimental results show that a training data set with more realistic reflection features and multiple sampling rates can enrich the data set variations in the structure and waveform signatures, thus significantly enhancing the fault segmentation. Moreover, a novel loss function we developed outperforms others with notable margins. Last, but not least, it is necessary to apply a test-time augmentation strategy by merging predictions with multiple scales and rotations in the reference step because the CNN does not preserve transformation invariance. Based on the studies, we optimally train a properly designed CNN and apply it to multiple field examples, where we obtain accurate, clean, continuous fault detections, and quantitatively evaluate them with manual interpretations.

Deep learning-based microstructure analysis of multi-component heterogeneous composites during preparation

Monitoring microstructure evolution during the preparation has always been a difficult problem in the modification studies of SiC composite matrix. Here, we used X-ray tomography microscopy to observe the microstructure of SiCf/SiC-W-ZrB2 composites at different fabrication stage. Based on deep learning, the tracking of the densification process of matrix-modified SiCf/SiC composites was achieved and its suitability for microstructure reconstruction was also verified. The results showed that the average errors of reconstructed SiCf/SiC, pore and Metal (W/ZrB2) are respectively 7.53%, 8.31% and 0.96% by comparison with the segmentation results. Compared with the experimental results, the average error and the average relative error of reconstructed SiCf/SiC is less than 3% and 3.74%.

Beyond pixel: Superpixel-based MRI segmentation through traditional machine learning and graph convolutional network

Background and Objective:Tendon segmentation is crucial for studying tendon-related pathologies like tendinopathy, tendinosis, etc. This step further enables detailed analysis of specific tendon regions using automated or semi-automated methods. This study specifically aims at the segmentation of Achilles tendon, the largest tendon in the human body. Methods:This study proposes a comprehensive end-to-end tendon segmentation module composed of a preliminary superpixel-based coarse segmentation preceding the final segmentation task. The final segmentation results are obtained through two distinct approaches. In the first approach, the coarsely generated superpixels are subjected to classification using Random Forest (RF) and Support Vector Machine (SVM) classifiers to classify whether each superpixel belongs to a tendon class or not (resulting in tendon segmentation). In the second approach, the arrangements of superpixels are converted to graphs instead of being treated as conventional image grids. This classification process uses a graph-based convolutional network (GCN) to determine whether each superpixel corresponds to a tendon class or not. Results:All experiments are conducted on a custom-made ankle MRI dataset. The dataset comprises 76 subjects and is divided into two sets: one for training (Dataset 1, trained and evaluated using leave-one-group-out cross-validation) and the other as unseen test data (Dataset 2). Using our first approach, the final test AUC (Area Under the ROC Curve) scores using RF and SVM classifiers on the test data (Dataset 2) are 0.992 and 0.987, respectively, with sensitivities of 0.904 and 0.966. On the other hand, using our second approach (GCN-based node classification), the AUC score for the test set is 0.933 with a sensitivity of 0.899. Conclusions:Our proposed pipeline demonstrates the efficacy of employing superpixel generation as a coarse segmentation technique for the final tendon segmentation. Whether utilizing RF, SVM-based superpixel classification, or GCN-based classification for tendon segmentation, our system consistently achieves commendable AUC scores, especially the non-graph-based approach. Given the limited dataset, our graph-based method did not perform as well as non-graph-based superpixel classifications; however, the results obtained provide valuable insights into how well the models can distinguish between tendons and non-tendons. This opens up opportunities for further exploration and improvement.

Explicit-implicit priori knowledge-based diffusion model for generative medical image segmentation

The diffusion probabilistic model (DPM) has achieved unparalleled results in current image generation tasks, and some recent research works employed it in several computer vision tasks, such as image super-resolution, object detection, etc. Thanks to DPM's superior ability to generate fine-grained details, these research efforts have yielded significant successes. In this paper, we propose a new DPM-based generative medical image segmentation method, named EIDiffuSeg. Specifically, we first construct an explicit-implicit aggregation priori knowledge with directional supervision ability by mining the semantic distribution pattern in the frequency and spatial domains. Then, the explicit-implicit aggregation priori knowledge is integrated into the different encoding stages of the denoising backbone network using a novel unsupervised priori knowledge induction strategy, which can guide the model to generate a segmentation mask of the region of interest directionally from a random inference process. We evaluate our method on three medical image segmentation benchmark datasets with different modalities and achieve the best segmentation results compared to state-of-the-art methods. Especially, compared to several current diffusion-based image segmentation methods, we achieved a 9% Dice improvement in the polyp segmentation benchmark. Our code will be available at https://github.com/Notmezhan/EIDiffuSeg.

A Comparative Analysis of U-Net and Vision Transformer Architectures in Semi-Supervised Prostate Zonal Segmentation.

The precise segmentation of different regions of the prostate is crucial in the diagnosis and treatment of prostate-related diseases. However, the scarcity of labeled prostate data poses a challenge for the accurate segmentation of its different regions. We perform the segmentation of different regions of the prostate using U-Net- and Vision Transformer (ViT)-based architectures. We use five semi-supervised learning methods, including entropy minimization, cross pseudo-supervision, mean teacher, uncertainty-aware mean teacher (UAMT), and interpolation consistency training (ICT) to compare the results with the state-of-the-art prostate semi-supervised segmentation network uncertainty-aware temporal self-learning (UATS). The UAMT method improves the prostate segmentation accuracy and provides stable prostate region segmentation results. ICT plays a more stable role in the prostate region segmentation results, which provides strong support for the medical image segmentation task, and demonstrates the robustness of U-Net for medical image segmentation. UATS is still more applicable to the U-Net backbone and has a very significant effect on a positive prediction rate. However, the performance of ViT in combination with semi-supervision still requires further optimization. This comparative analysis applies various semi-supervised learning methods to prostate zonal segmentation. It guides future prostate segmentation developments and offers insights into utilizing limited labeled data in medical imaging.

A hierarchical progressive recognition network for building change detection in high‐resolution remote sensing images

AbstractBuilding change detection (BCD) plays a crucial role in urban planning and development. However, several pressing issues remain unresolved in this field, including false detections of buildings in complex backgrounds, the occurrence of jagged edges in segmentation results, and detection blind spots in densely built‐up areas. To address these challenges, this study innovatively proposes a Hierarchical Adaptive Gradual Recognition Network (HAGR‐Net) to improve the accuracy and robustness of BCD. Additionally, this research is the first to employ the Reinforcement Learning Optimization Algorithm Based on Particle Swarm (ROPS) to optimize the training process of HAGR‐Net, thereby accelerating the training process and reducing memory overhead. Experimental results indicate that the optimized HAGR‐Net outperforms state‐of‐the‐art methods on the WHU_CD, Google_CD, and LEVIR_CD data sets, achieving F1 scores of 93.13%, 85.31%, and 91.72%, and mean intersection over union (mIoU) scores of 91.20%, 85.99%, and 90.01%, respectively.

Applying a mask R-CNN machine learning algorithm for segmenting electron microscope images of ceramic bronze-casting moulds

Material characteristics of casting moulds are crucial for understanding the evolution and diversification of bronze ritual vessel production in Bronze Age China. During relevant studies, a Back Scattered Electron (BSE) image detector is commonly employed to analyze mould microstructure, effectively revealing the volume ratios and shape features of the clay matrix, silt/sand particles, and voids. It is always challenging to analyze and cross-compare these BSE images quantitatively since they typically contain numerous phases with highly irregular shapes. Traditionally, time consuming manual point counting or multi-step image processing were used to obtain semi-quantitative results. Addressing these challenges, we have proposed a deep learning method called BCM-SegNet, an optimized Mask R-CNN-based algorithm for segmenting BSE images of bronze casting moulds and cores. Using the proposed method, key parameters, such as area, Feret diameter, roundness, and solidity of segmented particles, can be provided based on well segmented results, even for the images with complex background. Experimental outcomes show that the algorithm achieves a segmentation precision of 95% and an accuracy of around 91%, demonstrating its strong generalization capability. This study provides a significant foundation for micro-feature analysis of archaeological ceramic materials, classification of particles, and determination of technological processes in archaeological research.

Cross-modal Transfer Learning Based on an Improved CycleGAN Model for Accurate Kidney Segmentation in Ultrasound Images

ObjectiveDeep-learning algorithms have been widely applied in the field of automatic kidney ultrasound (US) image segmentation. However, obtaining a large number of accurate kidney labels clinically is very difficult and time-consuming. To solve this problem, we have proposed an efficient cross-modal transfer learning method to improve the performance of the segmentation network on a limited labeled kidney US dataset. MethodsWe aim to implement an improved image-to-image translation network called Seg-CycleGAN to generate accurate annotated kidney US data from labeled abdomen computed tomography images. The Seg-CycleGAN framework primarily consists of two structures: (i) a standard CycleGAN network to visually simulate kidney US from a publicly available labeled abdomen computed tomography dataset; (ii) and a segmentation network to ensure accurate kidney anatomical structures in US images. Based on the large number of simulated kidney US images and small number of real annotated kidney US images, we then aimed to employ a fine-tuning strategy to obtain better segmentation results. ResultsTo validate the effectiveness of the proposed method, we tested this method on both normal and abnormal kidney US images. The experimental results showed that the proposed method achieved a segmentation accuracy of 0.8548 in dice similarity coefficient on all testing datasets and 0.7622 on the abnormal testing dataset. ConclusionsCompared with existing data augmentation and transfer learning methods, the proposed method improved the accuracy and generalization of the kidney US image segmentation network on a limited number of training datasets. It therefore has the potential to significantly reduce annotation costs in clinical settings.

Bubble behavior parameters extraction and analysis during pool boiling based on deep-learning method

The nucleate pool boiling plays an important role in thermal and chemical engineering applications. Analyzing bubble dynamics at nucleation site is crucial to improve the understanding of boiling heat transfer mechanism. Quantitative extraction of bubble parameters from high-speed visualized images is a labor-intensitive and time-consuming task making it necessary for automatically detect single bubble growth and measure boiling characteristic parameters.In the present work, we proposed a deep learning based self-adaptive statistical algorithm for extraction of bubble behavior parameters quickly and automatically from numerous high-speed visualization images looking from the side view of a boiling chamber. A dataset was constructed for training and performance evaluation based on experimental data of saline solution pool boiling. The StarDist and U-Net convolutional neural network were combined in the algorithm so that more exact segmentation of the bubbles can be identified. Based on the segmentation results, a post-processing program was developed to extract the sequential variation of bubbles during consecutive cycles at nucleation sites. The dynamic characteristic parameters that affect heat transfer, such as nucleation density, bubble departure diameter, departure frequency, and wait time under different heat flux were obtained quantitatively. The comparison of automatic extraction algorithm and manual processing proves the reliability and superiority of our method. This work indicates that the proposed method has great potential to be widely applied as an efficient and universal tool for processing different types of bubble shadowgraph images.

VSmTrans: A hybrid paradigm integrating self-attention and convolution for 3D medical image segmentation

PurposeVision Transformers recently achieved a competitive performance compared with CNNs due to their excellent capability of learning global representation. However, there are two major challenges when applying them to 3D image segmentation: i) Because of the large size of 3D medical images, comprehensive global information is hard to capture due to the enormous computational costs. ii) Insufficient local inductive bias in Transformers affects the ability to segment detailed features such as ambiguous and subtly defined boundaries. Hence, to apply the Vision Transformer mechanism in the medical image segmentation field, the above challenges need to be overcome adequately. MethodsWe propose a hybrid paradigm, called Variable-Shape Mixed Transformer (VSmTrans), that integrates self-attention and convolution and can enjoy the benefits of free learning of both complex relationships from the self-attention mechanism and the local prior knowledge from convolution. Specifically, we designed a Variable-Shape self-attention mechanism, which can rapidly expand the receptive field without extra computing cost and achieve a good trade-off between global awareness and local details. In addition, the parallel convolution paradigm introduces strong local inductive bias to facilitate the ability to excavate details. Meanwhile, a pair of learnable parameters can automatically adjust the importance of the above two paradigms. Extensive experiments were conducted on two public medical image datasets with different modalities: the AMOS CT dataset and the BraTS2021 MRI dataset. ResultsOur method achieves the best average Dice scores of 88.3 % and 89.7 % on these datasets, which are superior to the previous state-of-the-art Swin Transformer-based and CNN-based architectures. A series of ablation experiments were also conducted to verify the efficiency of the proposed hybrid mechanism and the components and explore the effectiveness of those key parameters in VSmTrans. ConclusionsThe proposed hybrid Transformer-based backbone network for 3D medical image segmentation can tightly integrate self-attention and convolution to exploit the advantages of these two paradigms. The experimental results demonstrate our method's superiority compared to other state-of-the-art methods. The hybrid paradigm seems to be most appropriate to the medical image segmentation field. The ablation experiments also demonstrate that the proposed hybrid mechanism can effectively balance large receptive fields with local inductive biases, resulting in highly accurate segmentation results, especially in capturing details. Our code is available at https://github.com/qingze-bai/VSmTrans.

Sky-view factor enhanced doline delineation: A comparative methodological review based on case studies in Slovenia

This study presents a novel method for the accurate delineation of dolines in karst areas. The method is based on the use of hydrological tools and sky-view Factor in ArcGIS Pro and was implemented using a digital elevation model with a resolution of 1 m in four study areas in Slovenia. We manually delineated dolines at four test areas and compared them with the results of the new method, as well as the results of the most commonly used method of hydrological filling and the results of U-Net segmentation. We calculated the average deviation of the perimeter and the differences in the basic morphometric properties. The hydrological filling method cannot be accepted as a suitable method for doline delineation and should only be used for doline identification. The method based on the U-Net segmentation performed better, but the results contain landforms that are not enclosed depressions and therefore cannot be considered as dolines. The new method utilizes the advantages of hydrological filling and at the same time improves the doline delineation. We conclude that the presented method is the most suitable for automatic doline identification and delineation among the automatic methods discussed in this paper. This study further quantifies the enhancements achieved with the new method, highlighting the specific improvements in perimeter accuracy and the reliability of morphometric measurements.

Automatic Tumor Cellularity Measurement: AI-Based Pipeline for Multi-Organ Pathology Imaging.

Tumor Cellularity (TC) is an important metric for assessing organ tumor burden. However, manual cell counting is not feasible due to large volumes of pathology images and inconsistent measurements between pathologists. The PAIP 2023 Challenge aimed to solve this problem using AI. The challenge presented two main obstacles: the need to evaluate pancreas-trained data for effective use in the colon, and the common miscounting of clustered cells in segmentation results. To address these, we proposed a novel pipeline. It included channel normalization, which standardizes RGB values to ensure consistent model performance across different organs. By introducing CacoX, a specialized model for accurate cell segmentation, we used Coordinate Attention Gates for accurate cell localization and non-local learning. Finally, the implementation of a watershed algorithm allowed the automatic separation of clustered cells. This approach secured 3rd place in the PAIP 2023 Challenge with an impressive ICC score of 95.69%.

FastUNet: Fast hierarchical multi-patch underwater enhancement network for industrial recirculating aquaculture

The industrialized aquaculture with recirculating water systems has gained increasing attention and rapid development in recent years. However, the existing underwater enhancement methods are time-consuming and severely hinder their online application in industrialized aquaculture due to the requirement for real-time processing. We found that their main limitations are the insufficient utilization of the hierarchical network approach and valuable prior information in underwater scenarios. To address this problem, we propose a novel architecture called FastUNet and introduce a visual enhancement benchmark dataset specifically designed for industrialized aquaculture with recirculating water systems. Surprisingly, by learning the fast multi-patch hierarchical module and employing prior constraints including color loss and homology loss, FastUNet achieves comparable or even better results than many deep learning methods. In terms of runtime, it is up to 30 times faster than current underwater enhancement methods. The processing time for images with a resolution of 5474 × 3653 is only 0.137s. The real application scenario of factory recirculating aquaculture is considered by FastUNet, and the embedding of prior information under the fast network framework is explored. In this paper, 14 state-of-the-art underwater enhancement methods are compared, and excellent results are obtained on 2 Full-reference datasets and 2 No-referenced datasets. The application test results of SIFT significant point detection, geometric rotation, edge detection, target segmentation and target detection were also presented, which strongly verified the feasibility of FastUNet in the field of factory recirculating aquaculture.

An Enhanced Approach to Intelligent Computer-Assisted Localization of Liver Tumor on Computed Tomography Images

Computed Tomography (CT) Imaging is frequently used to find liver cancer. CT imaging of the liver generates cross-sectional images of the abdominal region. The task of segmenting a liver tumor on CT images is tedious due to the anatomic complexity of the liver. Therefore, liver tumor detection has been performed manually or semi-automatically which is an exhaustive process. Further, a tumor may be present on only some slices of the CT scan and hence there is a need to automatically identify images that contain tumors from acquired CT images and then locate a region of the tumor on those images for further diagnosis. Automatic detection helps the radiologist to obtain fast, accurate, and effective results. Accordingly, in this paper, an enhanced approach is proposed based on deep learning, efficient fuzzy c-means (EFCM) clustering, and a region-based algorithm for the detection of a liver tumor. A deep learning model built on a convolutional neural network is used to identify whether the CT image contains a tumor. If it contains the tumor, then the proposed EFCM and region-based algorithms are applied to locate the tumor on the CT image. The EFCM algorithm is combined with a region-based technique to automatically determine the initial pixel and threshold values required to segment the liver tumor. The proposed method is applied to several CT abdominal images. The results of the tumor segmentation are validated by comparing them with the tumor regions detected by radiologists. The estimated results of the experiment show that the proposed technique successfully detects the liver tumor on the CT image with less than 2% relative error.

A Novel Classification Approach for Retinal Disease Using Improved Gannet Optimization‐Based Capsule DenseNet

ABSTRACTAn unusual condition of the eye called diabetic retinopathy affects the human retina and is brought on by the blood's constant rise in insulin levels. Loss of vision is the result. Diabetic retinopathy can be improved by receiving an early diagnosis to prevent further damage. A cost‐effective method of accumulating medical treatments is through appropriate DR screening. In this work, deep learning framework is introduced for the accurate classification of retinal diseases. The proposed method processes retinal fundus images obtained from databases, addressing noise and artifacts through an improved median filter (ImMF). It leverages the UNet++ model for precise segmentation of the disease‐affected regions. UNet++ enhances feature extraction through cross‐stage connections, improving segmentation results. The segmented images are then fed as input to the improved gannet optimization‐based capsule DenseNet (IG‐CDNet) for retinal disease classification. The hybrid capsule DenseNet (CDNet) classifies disease and is optimized using the improved gannet optimization algorithm to boost classification accuracy. Finally, the accuracy and dice score values achieved are 0.9917 and 0.9652 on the APTOS‐2019 dataset.

Prototype-wise self-knowledge distillation for few-shot segmentation

Few-shot segmentation was proposed to obtain segmentation results for a image with an unseen class by referring to a few labeled samples. However, due to the limited number of samples, many few-shot segmentation models suffer from poor generalization. Prototypical network-based few-shot segmentation still has issues with spatial inconsistency and prototype bias. Since the target class has different appearance in each image, some specific features in the prototypes generated from the support image and its mask do not accurately reflect the generalized features of the target class. To address the support prototype consistency issue, we put forward two modules: Data Augmentation Self-knowledge Distillation (DASKD) and Prototype-wise Regularization (PWR). The DASKD module focuses on enhancing spatial consistency by using data augmentation and self-knowledge distillation. Self-knowledge distillation helps the model acquire generalized features of the target class and learn hidden knowledge from the support images. The PWR module focuses on obtaining a more representative support prototype by conducting prototype-level loss to obtain support prototypes closer to the category center. Broad evaluation experiments on PASCAL-5i and COCO-20i demonstrate that our model outperforms the prior works on few-shot segmentation. Our approach surpasses the state of the art by 7.5% in PASCAL-5i and 4.2% in COCO-20i.