Image Regions Research Articles

UAV rice panicle blast detection based on enhanced feature representation and optimized attention mechanism.

Rice blast is one of the most destructive diseases in rice cultivation, significantly threatening global food security. Timely and precise detection of rice panicle blast is crucial for effective disease management and prevention of crop losses. This study introduces ConvGAM, a novel semantic segmentation model leveraging the ConvNeXt-Large backbone network and the Global Attention Mechanism (GAM). This design aims to enhance feature extraction and focus on critical image regions, addressing the challenges of detecting small and complex disease patterns in UAV-captured imagery. Furthermore, the model incorporates advanced loss functions to handle data imbalances effectively, supporting accurate classification across diverse disease severities. The ConvGAM model, leveraging the ConvNeXt-Large backbone network and the Global Attention Mechanism (GAM), achieves outstanding performance in feature extraction, crucial for detecting small and complex disease patterns. Quantitative evaluation demonstrates that the model achieves an overall accuracy of 91.4%, a mean IoU of 79%, and an F1 score of 82% on the test set. The incorporation of Focal Tversky Loss further enhances the model's ability to handle imbalanced datasets, improving detection accuracy for rare and severe disease categories. Correlation coefficient analysis across disease severity levels indicates high consistency between predictions and ground truth, with values ranging from 0.962 to 0.993. These results confirm the model's reliability and robustness, highlighting its effectiveness in rice panicle blast detection under challenging conditions. The ConvGAM model demonstrates strong qualitative advantages in detecting rice panicle blast disease. By integrating advanced feature extraction with the ConvNeXt-Large backbone and GAM, the model achieves precise detection and classification across varying disease severities. The use of Focal Tversky Loss ensures robustness against dataset imbalances, enabling accurate identification of rare disease categories. Despite these strengths, future efforts should focus on improving classification accuracy and adapting the model to diverse environmental conditions. Additionally, optimizing model parameters and exploring advanced data augmentation techniques could further enhance its detection capabilities and expand its applicability to broader agricultural scenarios.

Radiomic Fingerprinting of the Peritumoral Edema in Brain Tumors

Background/Objectives: Tumor interactions with their surrounding environment, particularly in the case of peritumoral edema, play a significant role in tumor behavior and progression. While most studies focus on the radiomic features of the tumor core, this work investigates whether peritumoral edema exhibits distinct radiomic fingerprints specific to glioma (GLI), meningioma (MEN), and metastasis (MET). By analyzing these patterns, we aim to deepen our understanding of the tumor microenvironment’s role in tumor development and progression. Methods: Radiomic features were extracted from peritumoral edema regions in T1-weighted (T1), post-gadolinium T1-weighted (T1-c), T2-weighted (T2), and T2 Fluid-Attenuated Inversion Recovery (T2-FLAIR) sequences. Three classification tasks using those features were then conducted: differentiating between Low-Grade Glioma (LGG) and High-Grade Glioma (HGG), distinguishing GLI from MET and MEN, and examining all four tumor types, i.e., LGG, HGG, MET, and MEN, to observe how tumor-specific signatures manifest in peritumoral edema. Model performance was assessed using balanced accuracy derived from 10-fold cross-validation. Results: The radiomic fingerprints specific to tumor types were more distinct in the peritumoral regions of T1-c images compared to other modalities. The best models, utilizing all features extracted from the peritumoral regions of T1-c images, achieved balanced accuracies of 0.86, 0.81, and 0.76 for the LGG-HGG, GLI-MET-MEN, and LGG-HGG-MET-MEN tasks, respectively. Conclusions: This study demonstrates that peritumoral edema, as characterized by radiomic features extracted from MRIs, contains fingerprints specific to tumor type, providing a non-invasive approach to understanding tumor-brain interactions. The results of this study hold the potential for predicting recurrence, distinguishing progression from pseudo-progression, and assessing treatment-induced changes, particularly in gliomas.

Assessment of different U-Net backbones in segmenting colorectal adenocarcinoma from H&E histopathology.

Adenocarcinoma, the most prevalent type of colorectal cancer, makes up roughly 95 % of all cases and is associated with a notably high mortality rate. Owing to the various risk factors which might include personal choices and habits or genetic factors, the risk of developing the cancer for every individual might vary. However, given the statistics, the rate of acquiring the disease is pretty high. Therefore, based on the need for early detection and diagnosis of the disease, there is a pressing demand for an automated system to accurately identify adenocarcinoma in the colorectal region by utilizing the concept of binary segmentation wherein two classes are employed to indicate the presence as well as the absence of the condition. To address this, the project explored several deep learning-based segmentation methods-such as U-Net, Attention U-Net, U-Net with ResNet50 backbone, U-Net with MobileNet-v2 backbone, U-Net with EfficientNetB0 backbone, and U-Net with DenseNet121 backbone-to segment adenocarcinoma regions in histopathological images of the colon and rectum, which are essentially the various U-Net backbones. The performance of each method was then compared to identify the most effective approach, and subsequently, it was found that the U-Net with DenseNet121 backbone and U-Net with ResNet50 backbone performed better than the rest of the models in terms of accuracy with its respective training accuracy scores being 93.81 % and 93.39 % while the testing accuracy scores were 90.21 % and 89.81 %, respectively.

Image Navigation System for Thoracoscopic Surgeries Driven by Nuclear Medicine Utilizing Channel R-CNN

Breast cancer, a prevalent and significant cause of cancer-related mortality in women, often necessitates precise detection through nuclear medicine techniques. Despite the utility of computer-aided navigation in thoracoscopic surgeries like mastectomy, challenges persist in accurately locating and tracking target tissues amidst intricate surgical scenarios. This study introduces a novel system employing a channel R-CNN model to automatically segment target regions in thoracoscopic images and provide precise cutting curve indications for surgeons. By integrating a Detection Network Head and Thorax Network Head, this multi-channel framework outperforms existing single-task models, marking a pioneering effort in cutting curve indication for thoracoscopic procedures. Utilizing a specialized dataset, the model achieves a notable region segmentation mIOU of 79.4% and OPA of 83.2%. In cutting path planning, it attains an mIOU of 68.6% and OPA of 77.5%. The system operates at an average speed of 23.6 frames per second in videos, meeting the real-time response needs of surgical navigation systems. This research underscores the potential of advanced imaging and AI-driven solutions in enhancing precision and efficacy in thoracoscopic surgeries.

Three-Dimensional Reconstruction, Phenotypic Traits Extraction, and Yield Estimation of Shiitake Mushrooms Based on Structure from Motion and Multi-View Stereo

Phenotypic traits of fungi and their automated extraction are crucial for evaluating genetic diversity, breeding new varieties, and estimating yield. However, research on the high-throughput, rapid, and non-destructive extraction of fungal phenotypic traits using 3D point clouds remains limited. In this study, a smart phone is used to capture multi-view images of shiitake mushrooms (Lentinula edodes) from three different heights and angles, employing the YOLOv8x model to segment the primary image regions. The segmented images were reconstructed in 3D using Structure from Motion (SfM) and Multi-View Stereo (MVS). To automatically segment individual mushroom instances, we developed a CP-PointNet++ network integrated with clustering methods, achieving an overall accuracy (OA) of 97.45% in segmentation. The computed phenotype correlated strongly with manual measurements, yielding R2 > 0.8 and nRMSE < 0.09 for the pileus transverse and longitudinal diameters, R2 = 0.53 and RMSE = 3.26 mm for the pileus height, R2 = 0.79 and nRMSE = 0.12 for stipe diameter, and R2 = 0.65 and RMSE = 4.98 mm for the stipe height. Using these parameters, yield estimation was performed using PLSR, SVR, RF, and GRNN machine learning models, with GRNN demonstrating superior performance (R2 = 0.91). This approach was also adaptable for extracting phenotypic traits of other fungi, providing valuable support for fungal breeding initiatives.

No-reference image quality assessment based on information entropy vision transformer

ABSTRACT Convolutional Neural Networks and Transformers have good feature extraction but struggle in no-reference image quality assessment (NR - IQA) with real-world distortions. Vision Transformers, with self-attention, capture global and multi-scale features well. Local information entropy quantifies image region complexity, aligning with human perception. We propose IE - ViT, an NR - IQA model. It uses ViT for feature extraction, a dual-branch for multi-scale assessment, local information entropy for better human-perception simulation, and a global correlation loss. Experiments show it outperforms others on multiple datasets, with high scores on KADID - 10K.

Semantic-Guided Transformer Network for Crop Classification in Hyperspectral Images

The hyperspectral remote sensing images of agricultural crops contain rich spectral information, which can provide important details about crop growth status, diseases, and pests. However, existing crop classification methods face several key limitations when processing hyperspectral remote sensing images, primarily in the following aspects. First, the complex background in the images. Various elements in the background may have similar spectral characteristics to the crops, and this spectral similarity makes the classification model susceptible to background interference, thus reducing classification accuracy. Second, the differences in crop scales increase the difficulty of feature extraction. In different image regions, the scale of crops can vary significantly, and traditional classification methods often struggle to effectively capture this information. Additionally, due to the limitations of spectral information, especially under multi-scale variation backgrounds, the extraction of crop information becomes even more challenging, leading to instability in the classification results. To address these issues, a semantic-guided transformer network (SGTN) is proposed, which aims to effectively overcome the limitations of these deep learning methods and improve crop classification accuracy and robustness. First, a multi-scale spatial–spectral information extraction (MSIE) module is designed that effectively handle the variations of crops at different scales in the image, thereby extracting richer and more accurate features, and reducing the impact of scale changes. Second, a semantic-guided attention (SGA) module is proposed, which enhances the model’s sensitivity to crop semantic information, further reducing background interference and improving the accuracy of crop area recognition. By combining the MSIE and SGA modules, the SGTN can focus on the semantic features of crops at multiple scales, thus generating more accurate classification results. Finally, a two-stage feature extraction structure is employed to further optimize the extraction of crop semantic features and enhance classification accuracy. The results show that on the Indian Pines, Pavia University, and Salinas benchmark datasets, the overall accuracies of the proposed model are 98.24%, 98.34%, and 97.89%, respectively. Compared with other methods, the model achieves better classification accuracy and generalization performance. In the future, the SGTN is expected to be applied to more agricultural remote sensing tasks, such as crop disease detection and yield prediction, providing more reliable technical support for precision agriculture and agricultural monitoring.

Flood change detection model based on an improved U-net network and multi-head attention mechanism

This work aims to improve the accuracy and efficiency of flood disaster monitoring, including monitoring before, during, and after the flood, to achieve accurate extraction of flood disaster change information. A modified U-Net network model, incorporating the Transformer multi-head attention mechanism (TM), is developed specifically for the characteristics of Synthetic Aperture Radar (SAR) images. By integrating the TM, the model effectively prioritizes image regions relevant to flood disasters. The model is trained on a substantial volume of annotated SAR image data, and its performance is assessed using metrics such as loss function, accuracy, and precision. Experimental findings demonstrate significant improvements in loss value, accuracy, and precision compared to existing models. Specifically, the accuracy of the model algorithm in this work reaches 95.52%, marking a 3.46% improvement over the baseline U-Net network. Additionally, the developed model achieves an accuracy of 90.11% while maintaining a loss value of approximately 0.59, whereas other model algorithms exceed a loss value of 0.74. Thus, this work not only introduces a novel technical approach for flood disaster monitoring but also has the potential to enhance disaster response procedures and provide scientific evidence for disaster management and risk assessment processes.

Pericoronary adipose tissue feature analysis in computed tomography calcium score images in comparison to coronary computed tomography angiography.

We investigated the feasibility and advantages of using non-contrast CT calcium score (CTCS) images to assess pericoronary adipose tissue (PCAT) and its association with major adverse cardiovascular events (MACE). PCAT features from coronary computed tomography angiography (CCTA) have been shown to be associated with cardiovascular risk but are potentially confounded by iodine. If PCAT in CTCS images can be similarly analyzed, it would avoid this issue and enable its inclusion in formal risk assessment from readily available, low-cost CTCS images. To identify coronaries in CTCS images that have subtle visual evidence of vessels, we registered CTCS with paired CCTA images having coronary labels. We developed an "axial-disk" method giving regions for analyzing PCAT features in three main coronary arteries. We analyzed hand-crafted and radiomic features using univariate and multivariate logistic regression prediction of MACE and compared results against those from CCTA. Registration accuracy was sufficient to enable the identification of PCAT regions in CTCS images. Motion or beam hardening artifacts were often prevalent in "high-contrast" CCTA but not CTCS. Mean HU and volume were increased in both CTCS and CCTA for the MACE group. There were significant positive correlations between some CTCS and CCTA features, suggesting that similar characteristics were obtained. Using hand-crafted/radiomics from CTCS and CCTA, AUCs were 0.83/0.79 and 0.83/0.77, respectively, whereas Agatston gave AUC = 0.73. Preliminarily, PCAT features can be assessed from three main coronary arteries in non-contrast CTCS images with performance characteristics that are at the very least comparable to CCTA.

CERVIXNET: An Efficient Approach for the Detection and Classifications of the Cervigram Images Using Modified Deep Learning Architecture.

The earlier detection of cervical cancer in women patients can save human life. This article proposes a novel methodology for detecting abnormal cervigram images from healthy cervigram images and segments the cancer regions in the abnormal cervigram images using the deep learning method. The conventional deep learning architecture has been modified into the proposed CervixNet architecture to improve the cervical cancer detection rate. This methodology is constituted of a training and testing process, where the training process generates the training sequences individually for healthy cervigram images and the cancer case cervigram images. The testing process tests the cervigram images into either a healthy or cancer cases using the training sequences generated through the training process. During the testing process of the proposed system, the cancer segmentation algorithm was applied on the abnormal cervigram image to detect and segment the pixels belonging to cancer. Finally, the performance has been carried out on the segmented cancer cervical images for the ground truth images. This proposed methodology has been evaluated on the cervigrams on IMODT and Guanacaste databases. Its performance has been analyzed concerning cancer pixel sensitivity, cancer pixel specificity and cancer pixel accuracy. This research work obtains 98.69% Cancer Pixel Sensitivity (CPS), 98.76% Cancer Pixel Specificity (CPSP), and 99.27% Cancer Pixel Accuracy (CPA) for the set of cervigram images in the IMODT database. This research work obtains 99.22% CPS, 99.03% CPSP, and 99.01% CPA for the set of cervigram images in Guanacaste database. These experimental results of the proposed work have been significantly compared with the state-of-the-art methods and show the significance and novelty of the proposed works.

Enhancing Blood Cell Diagnosis Using Hybrid Residual and Dual Block Transformer Network

Leukemia is a life-threatening blood cancer that affects a large cross-section of the population, which underscores the great need for timely, accurate, and efficient diagnostic solutions. Traditional methods are time-consuming, subject to human vulnerability, and do not always grasp the subtle morphological differences that form the basic discriminatory features among different leukemia subtypes. The proposed residual vision transformer (ResViT) model breaks these limitations by combining the advantages of ResNet-50 for high dimensional feature extraction and a vision transformer for global attention to the spatial features. ResViT can extract low-level features like texture and edges as well as high-level features like patterns and shapes from the leukemia cell images. Furthermore, we designed a dual-stream ViT with a convolution stream for local details and a transformer stream for capturing the global dependencies, which enables ResViT to pay attention to multiple image regions simultaneously. The evaluation results of the proposed model on the two datasets were more than 99%, which makes it an excellent candidate for clinical diagnostics.

Neural Radiance Fields for High-Fidelity Soft Tissue Reconstruction in Endoscopy.

The advancement of neural radiance fields (NeRFs) has facilitated the high-quality 3D reconstruction of complex scenes. However, for most NeRFs, reconstructing 3D tissues from endoscopy images poses significant challenges due to the occlusion of soft tissue regions by invalid pixels, deformations in soft tissue, and poor image quality, which severely limits their application in endoscopic scenarios. To address the above issues, we propose a novel framework to reconstruct high-fidelity soft tissue scenes from low-quality endoscopic images. We first construct an EndoTissue dataset of soft tissue regions in endoscopic images and fine-tune the Segment Anything Model (SAM) based on EndoTissue to obtain a potent segmentation network. Given a sequence of monocular endoscopic images, this segmentation network can quickly obtain the tissue mask images. Additionally, we incorporate tissue masks into a dynamic scene reconstruction method called Tensor4D to effectively guide the reconstruction of 3D deformable soft tissues. Finally, we propose adopting the image enhancement model EDAU-Net to improve the quality of the rendered views. The experimental results show that our method can effectively focus on the soft tissue regions in the image, achieving higher fidelity in detail and geometric structural integrity in reconstruction compared to state-of-the-art algorithms. Feedback from the user study indicates high participant scores for our method.

Identification of the rank and kind of ten coals including bituminous and anthracite based on near-infrared spectroscopy and interpretable GASF-ResNet18 method

ABSTRACT Different ranks and kinds of coal have different properties (volatile, ash, etc.), which affect the process parameters in clean utilization of coal. Based on near-infrared (NIR) spectroscopy data, the 1D spectral data of 670 coal samples were encoded into 2D images with Gramian Angular Summation Field (GASF) method as input data, and a convolutional neural network (CNN) known as ResNet18 was applied to identify the rank and kind of coal. On a dataset containing seven kinds of bituminous coal and three kinds of anthracite coal, the proposed GASF-ResNet18 method achieves an accuracy of 96.3% for coal rank identification, and 94.8% for coal kind identification, even in the presence of moisture interference. Compared with traditional machine learning methods (partial least squares discriminant analysis, back propagation network, logistic regression, and support vector machine), GASF-ResNet18 method has a higher accuracy. It was attributed to ResNet18 catching the key information provided by the following GASF image regions: the combination of the absorption rate in 1250–1700 nm in column and 1550–1700 nm in row, and the combination of 900–1150 nm in column and 1150–1450 nm in row. The results showed GASF-ResNet18 method based on NIR data can be used to rapidly identify the rank and kind of coal.

Endoscopic ultrasonography-based intratumoral and peritumoral machine learning ultrasomics model for predicting the pathological grading of pancreatic neuroendocrine tumors

ObjectivesThe objective is to develop and validate intratumoral and peritumoral ultrasomics models utilizing endoscopic ultrasonography (EUS) to predict pathological grading in pancreatic neuroendocrine tumors (PNETs).MethodsEighty-one patients, including 51 with grade 1 PNETs and 30 with grade 2/3 PNETs, were included in this retrospective study after confirmation through pathological examination. The patients were randomly allocated to the training or test group in a 6:4 ratio. Univariate and multivariate logistic regression were used for screening clinical and ultrasonic characteristics. Ultrasomics is ultrasound-based radiomics. Ultrasomics features were extracted from both the intratumoral and peritumoral regions of conventional EUS images. Subsequently, the dimensionality of these radiomics features was reduced using the least absolute shrinkage and selection operator (LASSO) algorithm. A machine learning algorithm, namely multilayer perception (MLP), was employed to construct prediction models using only the nonzero coefficient features and retained clinical features, respectively.ResultsOne hundred seven ultrasomics features based on EUS were extracted, and only features with nonzero coefficients were ultimately retained. Among all the models, the combined ultrasomics model achieved the greatest performance, with an AUC of 0.858 (95% CI, 0.7512 - 0.9642) in the training group and 0.842 (95% CI, 0.7061 - 0.9785) in the test group. A calibration curve and a decision curve analysis (DCA) also demonstrated its accuracy and utility.ConclusionsThe integrated model using EUS ultrasomics features from intratumoral and peritumoral tumors accurately predicts PNETs' pathological grades pre-surgery, aiding personalized treatment planning.Trial registrationChiCTR2400091906.

Development and implementation of a deep learning-based smart evaluation system for animated films

Animation has a rich history of over a century, and scholarly interest in the critique of animated films has evolved alongside its development. With the advent of the Internet and advancements in computer technology, traditional methods of animated movie appreciation are increasingly challenged by new complexities. This paper addresses these challenges by presenting an advanced intelligent system for evaluating animated films, utilizing cutting-edge deep learning models. Our approach begins with a pre-trained deep learning model to classify image categories and extract scene characteristics, while a primary color extraction algorithm captures key color features. Building on this, we propose an innovative deep convolutional neural network (CNN) augmented with fully connected layers to integrate multimodal features, enhancing the accuracy of film evaluations. Furthermore, we introduce a novel deep learning framework designed to detect pixel regions in images that evoke emotional responses. By employing visualization techniques within the convolutional network, we identify emotionally significant regions and enhance these features during initial training, improving the system’s emotional sensitivity. Experimental results demonstrate that our system not only improves the accuracy and efficiency of animated film evaluations but also surpasses existing methods, offering a more nuanced, emotion-aware approach to animation critique. This work advances the field by providing a robust, automated tool for evaluating animated films, offering deeper insights into their emotional and visual components, and contributing to the ongoing evolution of AI-driven film appreciation systems.

A Heuristic Strategy Assisted Deep Learning Models for Brain Tumor Classification and Abnormality Segmentation

ABSTRACTBrain tumors are prevalent forms of malignant neoplasms that, depending on their type, location, and grade, can significantly reduce life expectancy due to their invasive nature and potential for rapid progression. Accordingly, brain tumors classification is an essential step that allows doctors to perform appropriate treatment. Many studies have been done in the sector of medical image processing by employing computational methods to effectively segment and classify tumors. However, the larger amount of information collected by healthcare images prohibits the manual segmentation process in a reasonable time frame, reducing error measures in healthcare settings. Therefore, automated and efficient techniques for segmentation are crucial. In addition, various visual information, noisy images, occlusion, uneven image textures, confused objects, and other features may impact the process. Therefore, the implementation of deep learning provides remarkable results in medicinal image processing, particularly in the segmentation and classification process. However, conventional deep learning‐assisted methods struggle with complex structures and dimensional issues. Thus, this paper develops an effective technique for diagnosing brain tumors. The main aspect of the proposed system is to classify the brain tumor types by segmenting the affected regions of the raw images. This novel approach can be applied for various applications like diagnostic centers, decision‐making tools, clinical trials, medical research institutes, disease prognosis, and so on. Initially, the requisite images are collected from standard datasets and further, it is subjected to the segmentation period. In this stage, the Multi‐scale and Dilated TransUNet++ (MDTUNet++) model is employed to segment the abnormalities. Further, the segmented images are given into an Adaptive Dilated Dense Residual Attention Network (ADDRAN) to classify the brain tumor types. Here, to optimize the ADDRAN technique's parameters, an Improved Hermit Crab Optimizer (IHCO) is supported, which increases the accuracy rates of the overall network. Finally, the numerical examination is conducted to guarantee the robustness and usefulness of the designed model by contrasting it with other related techniques. For Dataset 1, the accuracy value attains 93.71 for the proposed work compared to 87.86 for CNN, 90.18 for DenseNet, and 89.56 and 90.96 for RAN and DRAN, respectively. Thus, supremacy has been achieved for the recommended system while detecting the brain tumor types.

Depth Estimation Based on MMwave Radar and Camera Fusion with Attention Mechanisms and Multi-Scale Features for Autonomous Driving Vehicles

Autonomous driving vehicles have strong path planning and obstacle avoidance capabilities, which provide great support to avoid traffic accidents. Autonomous driving has become a research hotspot worldwide. Depth estimation is a key technology in autonomous driving as it provides an important basis for accurately detecting traffic objects and avoiding collisions in advance. However, the current difficulties in depth estimation include insufficient estimation accuracy, difficulty in acquiring depth information using monocular vision, and an important challenge of fusing multiple sensors for depth estimation. To enhance depth estimation performance in complex traffic environments, this study proposes a depth estimation method in which point clouds and images obtained from MMwave radar and cameras are fused. Firstly, a residual network is established to extract the multi-scale features of the MMwave radar point clouds and the corresponding image obtained simultaneously from the same location. Correlations between the radar points and the image are established by fusing the extracted multi-scale features. A semi-dense depth estimation is achieved by assigning the depth value of the radar point to the most relevant image region. Secondly, a bidirectional feature fusion structure with additional fusion branches is designed to enhance the richness of the feature information. The information loss during the feature fusion process is reduced, and the robustness of the model is enhanced. Finally, parallel channel and position attention mechanisms are used to enhance the feature representation of the key areas in the fused feature map, the interference of irrelevant areas is suppressed, and the depth estimation accuracy is enhanced. The experimental results on the public dataset nuScenes show that, compared with the baseline model, the proposed method reduces the average absolute error (MAE) by 4.7–6.3% and the root mean square error (RMSE) by 4.2–5.2%.

Staging of prostate Cancer with ultra-fast PSMA-PET scans enhanced by AI.

PSMA-PET is a reference standard examination for patients with prostate cancer, but even using recently introduced digital PET detectors image acquisition with standard field-of-view scanners is still in the range of 20min. This may cause limited access to examination slots because of the growing demand for PSMA-PET. Ultra-fast PSMA-PET may enhance throughput but comes at the cost of poor image quality. The aim of this manuscript is to evaluate the accuracy of AI-enhanced ultra-fast PSMA-PET for staging of patients with prostate cancer. A total number of 357 whole-body [68Ga]Ga-PSMA-11 PET datasets were included. Patients underwent two digital PET scans, one at standard and one at ultra-fast speed (table speed: 0.6-1.2mm/s vs. 50mm/s). A modified pix2pixHD generative adversarial network to enhance the ultra-fast images was trained with 286 datasets and evaluated with the remaining 71 datasets. The staging accuracy of ultra-fast PSMA-PET and AI-enhanced ultra-fast PET was compared with the reference standard PET separately for miTNM regions proposed by PROMISE V2.0. The AI-network significantly improved the visual image quality and detection rate in most miTNM regions compared with the non-enhanced image data (T: 69.6% vs. 43.5%, p < 0.05; N: 46.3% vs. 27.8%, p < 0.01; M1a 64.4% vs. 47.5%, p < 0.01; M1b: 85.7% vs. 72.1%, p < 0.01). However, improvement was not significant for the M1c category (42.9 vs. 28.6%, p > 0.05). Missed lesions had a smaller SUVmax and lesion size compared with detected lesions (exemplary for N: 9.5 vs. 26.5 SUVmax; 4 vs. 10mm). SUVmax values of lesions were significantly different in all miTNM regions between the ultra-fast and reference standard PET, but only in the T-region between the AI-enhanced and reference standard PET. The AI-based image enhancement improved image quality and region detection rates by a mean of 17.9%. As the sensitivity of synthetic PET for small and low-uptake lesions was limited, a potential clinical use case could be disease monitoring in patients with high tumor volume and PSMA uptake undergoing PSMA radioligand therapy. The improvement in detection rate of distant metastases was not significant. This indicates that more training data is needed to ensure robust results also for lesions that have lower appearance frequency. Future studies on accelerated PSMA-PET seem warranted.

Skin Lesion Classification Through Test Time Augmentation and Explainable Artificial Intelligence.

Despite significant advancements in the automatic classification of skin lesions using artificial intelligence (AI) algorithms, skepticism among physicians persists. This reluctance is primarily due to the lack of transparency and explainability inherent in these models, which hinders their widespread acceptance in clinical settings. The primary objective of this study is to develop a highly accurate AI-based algorithm for skin lesion classification that also provides visual explanations to foster trust and confidence in these novel diagnostic tools. By improving transparency, the study seeks to contribute to earlier and more reliable diagnoses. Additionally, the research investigates the impact of Test Time Augmentation (TTA) on the performance of six Convolutional Neural Network (CNN) architectures, which include models from the EfficientNet, ResNet (Residual Network), and ResNeXt (an enhanced variant of ResNet) families. To improve the interpretability of the models' decision-making processes, techniques such as t-distributed Stochastic Neighbor Embedding (t-SNE) and Gradient-weighted Class Activation Mapping (Grad-CAM) are employed. t-SNE is utilized to visualize the high-dimensional latent features of the CNNs in a two-dimensional space, providing insights into how the models group different skin lesion classes. Grad-CAM is used to generate heatmaps that highlight the regions of input images that influence the model's predictions. Our findings reveal that Test Time Augmentation enhances the balanced multi-class accuracy of CNN models by up to 0.3%, achieving a balanced accuracy rate of 97.58% on the International Skin Imaging Collaboration (ISIC 2019) dataset. This performance is comparable to, or marginally better than, more complex approaches such as Vision Transformers (ViTs), demonstrating the efficacy of our methodology.

Efficient Image Inpainting for Handwritten Text Removal Using CycleGAN Framework

With the recent rise in the development of deep learning techniques, image inpainting—the process of restoring missing or corrupted regions in images—has witnessed significant advancements. Although state-of-the-art models are effective, they often fail to inpaint complex missing areas, especially when handwritten occlusions are present in the image. To address this issue, an image inpainting model based on a residual CycleGAN is proposed. The generator takes as input the image occluded by handwritten missing patches and generates a restored image, which the discriminator then compares with the original ground truth image to determine whether it is real or fake. An adversarial trade-off between the generator and discriminator motivates the model to improve its training and produce a superior reconstructed image. Extensive experiments and analyses confirm that the proposed method generates inpainted images with superior visual quality and outperforms state-of-the-art deep learning approaches.