Generative AI currently assists in solving various tasks in the general domain of computer vision. However, for precision medicine applications, deep domain knowledge is required to develop specialized tools for practical clinical tasks. Our study focuses on developing advanced computer vision approaches for critical histopathological tasks, particularly virtual staining of tissues, which accelerates the diagnostic process and provides pathologists with additional valuable information while reducing labor-intensive workflows. Different types of staining protocols, such as H&E (hematoxylin and eosin) and IHC (immunohistochemistry), provide complementary information by highlighting specific morphological on tissue slides. For precise analysis, pathologists may need to observe preliminary staining results to select the most appropriate diagnostic approach. However, current generative AI algorithms still face limitations in achieving sufficiently accurate results to address all clinical requirements. The primary challenges currently faced in this domain include: (1) limitations in available high-quality paired datasets that support reliable mapping between different staining types, (2) various artifacts that compromise image quality, and (3) insufficient alignment between multi-stain samples. To address these challenges, we investigate state-of-the-art techniques based on generative adversarial networks (GANs), vision transformers, and diffusion-based models. Prior to AI model training, we specifically address the critical task of co-registration of whole-slide images with different staining protocols to achieve more accurate alignment between paired samples, thereby minimizing mismatches and artifacts. When combined with advanced algorithms for biological structures and tumor area segmentation, these generative AI algorithms provide promising avenues for advancing both effective and precision medicine applications in histopathology.
Medical imaging datasets are often constrained by acquisition costs, scarce patient cohorts, and privacy restrictions. These limitations are particularly acute for subtle neuropathologies such as focal cortical dysplasia (FCD), where conventional MRI struggles to detect epileptogenic regions— posing a critical bottleneck for computer-aided diagnosis. Deep generative models offer a promising solution by synthesizing realistic data. Yet, in scenarios with limited and complex training data—such as 3D brain MRI-generated images often suffer from inadequate diversity. To overcome these limitations, we introduce a novel generative inpainting approach for synthesizing highquality medical images with realistic pathologies. Our approach trains a generative model on limited pathological data and leverages it to inpaint lesions into healthy MRI scans, preserving anatomical context while enhancing synthetic image diversity. We also introduce a dedicated mask-generation model to synthesize biologically plausible pathology masks, ensuring realistic lesion morphology and distribution across the brain. The addition of synthetic data into training dataset (neither validation or test) improves performance in lesion segmentation task. Both precision and Dice score increase up to 20%, which demonstrate the method's effectiveness in augmenting limited pathological data and enhancing FCD detection.
Transformer-based architectures have become the dominant approach for a wide array of machine learning tasks, including those in computer vision. Consequently, the prevalence of purely convolutional networks-particularly shallow-depth architectures for classification-has been in decline. In this work, we revisit Convolutional Neural Networks (CNNs) and propose a modern hybrid architecture that integrates Transformer-inspired components. Specifically, we introduce MLP Fusion, a model that incorporates Multi-Layer Perceptron (MLP) blocks, similar to those used in Vision Transformers, into CNN backbones prior to the classification stage. Additionally, we include intermediate $1 \times 1$ convolutional layers within the backbone. This fusion is intended to enhance the representational capacity of CNNs by enriching their embedding space. Experimental evaluations on the CIFAR-10 and CIFAR-100 datasets show that MLP Fusion achieves better performance compared to compact CNN models reported in the literature.
Interpersonal relationship recognition from images presents significant challenges due to the complex interplay of visual cues, contextual information, and social dynamics. This paper introduces a novel Dual-Graph Transformer architecture with Contextual-Support Nodes (CSNs) that addresses limitations in existing approaches. The proposed method employs parallel graph neural networks to process micro and macro contextual evidence separately, while introducing CSNs to ensure graph connectivity and compress heterogeneous features. Building upon progressive graph reasoning principles, the architecture incorporates linguistic descriptions alongside visual features, enhancing robustness to environmental uncertainties. A transformer-based fusion mechanism enables learnable inter-graph and intermodality information exchange without predefined rules. Experimental evaluation on the People in Social Context (PISC) dataset demonstrates superior performance compared to state-of-the-art methods, with significant improvements in relationship classification accuracy. The results highlight the importance of integrating extra-visual context and flexible fusion strategies in graph-based social relationship recognition systems.
The accurate detection of suspicious human actions, such as shoplifting, in surveillance videos is critical for ensuring public safety and enabling long-term crime prevention and crowd trust. However, this task remains challenging due to complex scene dynamics, varying camera viewpoints, and imbalanced data distributions. In this paper, we propose a new action recognition framework integrating fine-grained spatial localization of individuals with transformer-based temporal modeling to enhance person-level detection of suspicious behaviors in retail environments. Our approach leverages state-of-the-art object detection techniques to extract per-person regions of interest, which are then processed by a temporal encoder to effectively capture motion patterns indicative of shoplifting. Through a gap analysis, we identify that existing methods often fall short in handling diverse scene conditions and data imbalance—issues that our framework explicitly addresses. We evaluate our method on challenging datasets featuring varied store layouts and camera angles, achieving a sensitivity of 91.34% and an accuracy of 82.45%, surpassing the baseline by 33% and 21%, respectively. These results highlight the robustness and practical relevance of our approach for real-world retail surveillance applications requiring reliable and interpretable action recognition.
Magnetic resonance imaging (MRI) is essential for diagnosing acute ischemic stroke and assessing the extent of salvageable brain tissue, known as the ischemic penumbra. Perfusion-weighted imaging (PWI) provides crucial information for this assessment by allowing the calculation of the diffusionperfusion mismatch, a key factor in treatment choice. However, PWI is often unavailable in some clinical protocols due to logistical and technical constraints, such as administering an exogenous contrast agent. This work introduces a 3D conditional generative adversarial network (cGAN) designed to synthesize time-to-peak (TTP) perfusion maps from non-contrast MRI sequences. Our approach leverages spatial conditioning based on lesion annotations and employs a U-Net generator enhanced with attention mechanisms paired with a 3D PatchGAN discriminator. The quantitative evaluation shows that the model successfully generates high-fidelity TTP maps using apparent diffusion coefficient (ADC) and susceptibility-weighted imaging (SWI) as inputs, achieving structural similarity index measure (SSIM) scores as high as 0.89. These results demonstrate the promise of our method as a clinical decision-support tool and as an alternative for training models that estimate perfusion-related biomarkers, potentially overcoming the need for exogenous contrast in acute stroke evaluation.
Generative Adversarial Networks (GANs) offer a powerful framework for synthesizing post-disaster imagery, but their application to building damage assessment is limited by difficulties in generating realistic images of severely deformed structures, slow inference speeds that hinder real-time use in emergency scenarios, and a reliance on paired pre- and post-disaster data, which are rarely available in datasets collected by unmanned aerial vehicles (UAVs). To address these challenges, we propose Struct-CycleGAN, a cross-domain GAN model designed to generate high-fidelity and structurally consistent post-disaster building images from unpaired data. The model improves upon prior limitations in texture transformation and structural deformation synthesis while significantly enhancing inference speed for practical deployment. Compared to a commonly used state-of-the-art method, Struct-CycleGAN achieves nearly four times faster inference and obtains a perceptual quality score of 23.7 based on Fréchet Inception Distance (FID) evaluation. Further validation using a ResNet UNet segmentation framework shows that undamaged building areas in pre-disaster images decrease from 45% to 22% in the generated post-disaster outputs, demonstrating the model's effectiveness in simulating structural degradation. This work provides a scalable solution for synthesizing post-disaster building scenarios, supporting rapid damage assessment, emergency preparedness, and intelligent reconstruction efforts.
This paper introduces a framework for enhancing industrial safety through context-aware simulation. The system is built around a modular pipeline that captures real-time data, uses object detection (YOLOv8) and image captioning (BLIP) to interpret scenes, and communicates insights through middleware to a simulation awareness layer. This layer can inject hazards and coordinate responses. By providing richer contextual understanding, the system enables earlier and more effective safety interventions, leading to increased operator confidence compared to traditional threshold-based triggers. The approach demonstrates how advanced perception and language technologies can be integrated into simulation environments to support safer and more efficient industrial operations, while also revealing new behaviors that emerge from this deeper level of situational awareness.
Automated waste detection is essential for effective urban waste management. In this context, several recent studies based on machine learning have demonstrated strong potential in the classification and identification of waste. In this work, we present an in-depth study highlighting the evolution of YOLO (You Only Look Once) object detection models, from YOLOv5 to YOLOv12, applied to real-time waste classification. YOLO's speed, accuracy, and lightweight architecture make it an ideal solution for edge computing deployment. Evaluations conducted on the IURWDD (Indian Urban and Rural Waste Detection Dataset), which consists of annotated images representing various types of urban and rural waste, show that compact variants such as YOLOv11n and YOLOv12n achieve excellent accuracy with minimal latency. Through this study, we identified several major challenges, including class imbalance and domain variability, and proposed strategies such as transfer learning and lightweight optimization. Overall, YOLO-based models show strong potential for developing scalable, sustainable, and real-time waste management solutions that meet the demands of smart cities.
High-quality images are essential in medical ultrasound imaging to ensure reliable clinical diagnostics. Therefore, the quantification of ultrasound image quality is necessary to properly evaluate an ultrasound system. This paper provides an overview of the current methods used to assess ultrasound image quality. We explore two approaches: one based on quantifiable physical characteristics such as spatial resolution, signal-to-noise ratio, contrast, and depth of penetration, typically measured using ultrasound phantoms. The other approach is based on human visual perception of image quality. The latter includes objective computational algorithms classified as full-reference, by comparing a given image to a reference pristine image, and no-reference, that does not use a reference image. We generated a dataset of 4000 images of anechoic contrast lesions by simulation and evaluated ten different types of quality metrics. Finally, we discuss the relation between those metrics, select the most appropriate and suggest avenues for further investigation.
Motivation: In biological dosimetry a radiation dose is estimated using the average number of chromosomal aberrations per peripheral blood lymphocytes. This analysis is manually performed and better automating chromosomal aberration detection has become of paramount of importance to improve diagnosis reliability and reduce human expertise time. Results: We build an object detection model to automate chromosomal aberration detection using deep convolutional neural networks. We formulated the problem of rare aberration detection in metaphase images as a heatmap regression problem. We demonstrate significant performance improvements and provide statistical confidence intervals by using an ensemble of checkpoints that are selected at the end of each epoch during training and combined to generate a consensus model. We also propose in this work an approach to visualize training dynamics using low-dimension representation to better interpret the relationships between training stochasticity and ensemble diversity. This approach provided an conveniant explainable artificial intelligence tool to understand the mechanism of the chromosomal aberration detection of the model.
Endometrial cancer (EC) is one of the most prevalent gynecological malignancies, and fast, accurate diagnosis is critical for effective treatment. The current diagnostic workflow relies heavily on hysteroscopy followed by histopathological analysis, a process that is often time-consuming and prone to subjectivity. Accordingly, this study proposes a deep learning-based diagnostic support system designed to assist in post-hysteroscopic evaluation through automated image segmentation. The proposed solution incorporates three state-of-the-art architectures, namely EfficientNet, YOLOv11x, and Mask R-CNN, trained and tested on a dataset of hysteroscopic images annotated by experienced gynecologists. Comparative performance metric results show that Efficient-Net achieves precision of $\mathbf{9 2. 4 5 \%}$, recall of $\mathbf{9 3. 1 7 \%}$, F1-score of $\mathbf{9 2. 7 9 \%}$, and mAP50 of $\mathbf{9 7. 3 6 \%}$, outperforming other models across all metrics. Furthermore, an application is developed for potential clinical workflow integration of the models. It supports both image and video input, with features such as blur detection and confidence scoring. Usability testing with medical students indicates high user satisfaction and confirms the system's clinical potential. Overall, the findings support the use of segmentation-based AI tools to reduce diagnostic delays and improve objectivity in endometrial cancer detection.
Image fusion is a technique that integrates information from multiple sources into a single image to produce richer and more meaningful visual content. This study applied three classical fusion methods—Multi-Scale Transform (MST), Saliency-Based fusion, and Anisotropic Diffusion with KL Transform—to RGB and thermal image pairs from the LLVIP dataset. CLAHE preprocessing was applied to enhance contrast, and the resulting fusion images were evaluated in terms of object detection performance and image quality metrics. For object detection, YOLOv4, YOLOv4-Tiny, and YOLOv7 models were used. According to YOLOv4 results, MST fusion applied to CLAHE-preprocessed RGB images achieved 0.18% higher $\operatorname{mAP} {@} 0.5$ accuracy compared to thermal images alone and 9.53% higher accuracy compared to RGB images alone. Regarding the authors' knowledge, this is the first time the study presents the comparative analysis of these models on the LLVIP dataset. The fusion outputs were also evaluated using SSIM, Mutual Information, and $Q_{a b / F}$ metrics. The results indicate that the Saliency-Based fusion method yielded the best performance, particularly regarding visual quality.
State-of-the-art computer vision architectures frequently adopt U-Net architectures due to their flexibility and adaptability. However, their reliance on multi-dimensional convolutions introduces significant computational overhead, limiting practicality on low-power edge devices. We propose OneNet, a lightweight alternative that replaces 2D convolutions with channel-wise 1D convolutions in the encoder, paired with pixelunshuffle operations to maintain spatial awareness. Leveraging PixelShuffle-a technique known to enhance performance in super-resolution tasks-our design preserves segmentation accuracy while reducing parameter count by up to 47 %. We further explore a fully 1D encoder-decoder variant that achieves a 71 % model size reduction, with a trade-off in accuracy. Comprehensive evaluations across various mask-generation tasks demonstrate that OneNet matches U-Net performance in accuracy, making it a compelling choice for edge-based image segmentation. Moreover, the architecture's modular design allows easy adaptation to a broader range of convolutional applications. Code for the project is available at https://github.com/shbyun080/OneNet.
Automatic stitch counting remains a critical bottleneck for computer-aided knitwear design and quality control. We present the first fully automated pipeline that predicts the number of next stitches in crochet swatches using only a single RGB image. The task is framed as a three-class semantic segmentation problem (background, hole, and yarn). A lightweight U-Net model was trained on a curated dataset of 163 tightly-cropped swatch images, with 71 additional images reserved for validation. The system achieves accurate per-pixel classification and high agreement with ground-truth stitch counts. Training was performed on an NVIDIA A100 GPU via Google Colab. Our ablation studies highlight the importance of class-balanced loss design and targeted hole up-weighting for recall. The dataset, training code, and pre-trained weights will be made publicly available to foster reproducibility and follow-up research. Please contact the authors for exact performance benchmarks or system specifications.
Social media has become a vital source of real-time information in disaster response, yet the semantic inconsistency and noise in user-generated content present major challenges for automated analysis. This study introduces a multimodal filtering and classification framework to improve the reliability of disaster-related social media data. We evaluate two independent filtering strategies: the Tweet Metric Score, based on a linguistic heuristic score, and CLIPScore, a reference-free semantic alignment metric assessing image-text coherence. Using a curated dataset of crisis-related posts, we train multimodal classifiers that integrate CLIP/ResNet-50 and BERT features via intermediate fusion. Experimental results show that CLIPScore-filtered data consistently outperforms both unfiltered and heuristically filtered datasets, achieving $\mathbf{9 3. 4 7 \%}$ test accuracy and $\mathbf{9 3. 4 1 \%} \mathbf{F 1}$-score. These findings confirm that high-quality, semantically aligned data significantly enhances classification performance. Our approach highlights the importance of content alignment in multimodal crisis informatics and provides a scalable solution for improving situational awareness during emergencies.
Acute Lymphoblastic Leukemia (ALL) is an aggressive form of blood cancer characterized by the rapid proliferation of immature lymphoblasts in the bone marrow and peripheral blood. It is the most common type of leukemia in children and requires early, accurate, and rapid diagnosis to ensure effective treatment. However, conventional methods such as flow cytometry, while effective, are expensive, time-consuming, and heavily reliant on human expertise, limiting their accessibility. In this context, Artificial Intelligence (AI), particularly deep learning-based approaches, emerges as a promising solution for automating the screening of ALL from Peripheral Blood Smear (PBS) images. This study presents a comparative evaluation of several Convolutional Neural Network (CNN) architectures, including a lightweight, custom-designed model named SLIM-CNN, and vision transformer models. To simulate realworld clinical conditions, visual perturbations, such as blur and zoom were applied to the test set, enabling an assessment of the models' robustness. The results show that although CNNs including the custom SLIM-CNN model, perform well under ideal conditions, they remain sensitive to visual perturbations and lack robustness when faced with data variability. The lightweight SLIM-CNN specifically designed for automated ALL detection and classification, benefited from explainable AI (XAI) techniques, enabling visual identification of the model's attention regions. However, despite this interpretability, its stability remains limited under realistic image degradations. In contrast, vision transformer models demonstrate superior generalization capabilities and enhanced robustness in environments that simulate real-world clinical conditions, positioning them as strong candidates for effective integration into medical practice.
Accurate and timely assessment of the severity of the damage is vital for effective disaster response and prioritization of resources. However, the unstructured and noisy nature of multimodal social media data poses challenges for reliable automated analysis. This paper introduces a robust vision-language pipeline that combines domain-adapted caption generation with BLIP2, semantic alignment filtering using CLIPScore and ImageText Matching (ITM), and chain-of-thought reasoning with a large multimodal language model (LLM) to estimate quantitative severity scores. Evaluations of the CrisisMMD dataset demonstrate significant improvements over the baselines of traditional and unimodal fusion, achieving up to 80.2% precision and strong F1 performance for high-confidence severe damage cases. To address uncertain or low-confidence instances, we advocate a human-in-the-loop design that routes ambiguous cases for manual verification, ensuring trustworthy decision support for emergency response operations.
The complexity of training datasets which are characterized by a large number of classes with subtle inter-class differences, poses significant challenges for model efficiency and deployment on resource constrained environments. While large deep learning models excel at capturing these complex class structures, their computational demands make them impractical for deployment on resource-constrained devices. In contrast, smaller models, though more efficient, struggle to achieve comparable performance due to their limited capacity. Knowledge distillation helps bridge this gap by transferring knowledge from a large, powerful teacher model to a smaller, more efficient student model. In this paper, we introduce a specialized knowledge distillation method. Instead of using a single student model, we train multiple specialist models, each focusing on a subset of classes. This helps the models better distinguish between similar classes. Our approach automatically groups classes into meaningful clusters by averaging CLIP embeddings at the class level. We also utilize a cluster predictor that dynamically selects the most suitable specialist(s) for each input, balancing accuracy with efficiency. We test our method on ImageNet and show that these specialist models outperform traditional student models on their respective class subsets. Our approach provides a scalable way to build accurate and efficient models suitable for resource-constrained environments.
Agriculture is one of the main economic pursuits of the 64 Bangladeshi districts.: in Bangladesh, about seventy percent of the workforce rely on agriculture for their living. Bangladesh's gross national revenue is much enhanced by the country's rice farming but attacks of insect pests have a great impact on rice harvests. Different insect pests require different management measures, so the accurate identification of paddy field insects is a crucial task that allows, e.g., the application of the appropriate poison for specific insect pests. This will also prevent the wasteful use of ineffective insecticides. The main challenge addressed in this work is to detect and instantly segment small harmful insects in paddy fields. To address this problem, we have used deep convolutional neural network (DCNN) learning, based on Mask-RCNN, therefore enabling a technique for visual localisation and classification of agricultural pest insects. We have also developed our own dataset of harmful insect annotated images. In the proposed Mask-RCNN model we used a ResNet101 backbone, which can detect and segment at the same time. The proposed model achieves an AP@0.5 of 85.7%, a mAP of 63.8%, and an AR@10 of 68.5, therefore generating an anticipated accuracy of 75%. ResNet101 performs better on all measures. The suggested approach should be able to identify and classify the small harmful insects, with suitable accuracy, in real-world deployments.
