Computer Science and Mathematics

Abstract: 3D scene reconstruction and novel view synthesis have undergone a paradigm shift with the advent of 3D Gaussian Splatting (3DGS). Unlike computationally intensive volumetric rendering, 3DGS leverages a point-based representation with a differentiable tile-based rasterizer. This elegant synthesis achieves state-of-the-art visual fidelity at real-time rendering frame rates. As the 3DGS literature grows rapidly, existing surveys have primarily organized the field around downstream applications. In contrast, this survey provides a rasterization-centric analysis, treating the differentiable rasterization pipeline as the core computational engine of 3DGS. We begin by delineating the mathematical foundations of Gaussian splatting and tracing its lineage from classical volume rendering. Subsequently, we propose a fine-grained taxonomy that categorizes the literature across three hierarchical dimensions: (1) representation and optimization, (2) rasterization pipeline innovations, and (3) scenario-driven rasterization extensions. By deconstructing these algorithmic advances in rasterization, we offer quantitative insights into hardware-algorithm co-design and outline critical trajectories for future research. To support the community, we maintain a continually updated repository of relevant literature and open-source implementations at https://github.com/3DAgentWorld/Advanced3DGS.

Abstract: Accurate and up-to-date knowledge of land use and land cover represents one of the central challenges in spatial planning and landscape sciences. In this context, the present work introduces GANCIU (Geospatial Analysis with Neural Classification and Image Understanding), an original hybrid pipeline for the automatic extraction of man-made infrastructure from high-resolution satellite imagery. The primary methodological contribution lies in the sequential integration of four technologically heterogeneous components: a per-pixel Random Forest classifier, a guided image modulation step, edge detection via the Mumford-Shah variational functional solved through the Ambrosio-Tortorelli approximation, and final object delineation via the Segment Anything Model (SAM). Each component does not operate independently but conditions and informs the next: the RF probability map guides the modulation, which in turn directs the sensitivity of the variational step exclusively towards regions of interest; the AT edges provide spatial prompts to SAM, whose masks are finally filtered by the RF probability in an adaptive manner through a Gaussian mixture model. This progressive conditioning scheme constitutes the architectural core of GANCIU and distinguishes it from approaches that combine classification and segmentation in parallel or in purely sequential fashion without informational feedback between steps. The pipeline was applied to fourteen study areas located in Sardinia, a region characterized by a complex rural matrix and marked chromatic and geometric heterogeneity of surface types. The results reveal three distinct behavioral regimes as a function of scene difficulty: in scenes with clear radiometric contrast between infrastructure and agricultural background, the pipeline discriminates effectively at the first step; in scenes of intermediate difficulty, the AT+SAM+filter combination demonstrates a significant recovery capacity with respect to the uncertainty of the initial RF map; in scenes with dense agricultural texture, the current limitations of the training model emerge, with false positives on intensive crops representing the primary target for future development. A quantitative comparison with established methods, based on standard metrics including Precision, Recall, F1-score and Intersection over Union, is deferred to a subsequent evaluation that will extend the analysis to different geographical and sensor contexts.

Abstract: Lightweight vision models are often advanced through architecture-specific innovations or scaling strategies, which can make it difficult to distinguish generalizable design principles from specialized design choices. We hypothesize that lightweight architectures can be more effectively improved by systematically composing complementary components that provide distinct representational capabilities, rather than by optimizing individual building blocks in isolation. Motivated by this perspective, we propose EfficientMixer, a lightweight vision architecture that integrates efficient local feature extraction, spatial mixing, and lightweight global feature modeling within a unified component-level framework inspired by EfficientNet, ConvMixer, and MobileViT. This design enables structured interaction among complementary inductive biases under a fixed budget. Under matched budgets, EfficientMixer consistently outperforms strong lightweight baselines across multiple benchmarks while maintaining similar computational cost and demonstrating improved cross-domain generalization. Furthermore, EfficientMixer achieves larger performance gains under Self-Competitive Distillation, suggesting improved compatibility with advanced training strategies. Extensive ablation studies indicate that these improvements are largely associated with synergistic interactions among complementary architectural components rather than increased model parameters or architectural scaling. Overall, these results highlight component-level architectural composition as an effective and practical design principle for developing efficient and transferable lightweight vision models.

Abstract: Traditional Deep Learning (DL) models for single-image super-resolution (SISR) employ upsampling techniques such as interpolation, transpose convolution, or pixel-shuffle operations. These upsampling techniques are not explicitly aware of the frequency domain structure, leading to a loss of finer details. As a result, edges and structural high-information details are often not preserved. To address this limitation, we propose to introduce an the Inverse Discrete Wavelet Transform (IDWT) based upsampling approach, with a U-Net architecture as a case study. To this end, the encoder's downsampling part is replaced by the Discrete Wavelet Transform (DWT), and the decoder's upsampling part is replaced by the IDWT for ×2 single-image super-resolution. In the decoder's upsampling part, we propose a learned channel-splitting mechanism that partitions the feature map into the LH, HL, and HH sub-bands. The idea is to learn the high frequency sub-bands from the input by assuming the later to be an LL sub-band; the mechanism thus predicts the lost high-frequency wavelet information prior to IDWT reconstruction. The wavelet-based U-Net we propose utilizes DWT and IDWT to maintain high-frequency details, including edges and textures, thereby improving the quality of single-image super-resolution. Our model was trained on the DIV2K dataset, and all experiments were conducted under identical and controlled training settings across all baseline variants. Experimental results on natural-image and text-image benchmark datasets demonstrate that the proposed wavelet-based U-Net model achieves competitive and superior reconstruction performance for x2 super-resolution. In addition to the quantitative improvements in shape of PSNR and SSIM, visual comparisons further indicate that the proposed method reconstructs sharper and more detailed image structures. These findings suggest that integrating a wavelet transform into encoder-decoder networks helps preserve high-frequency details more effectively, providing an efficient and high-quality solution for single-image super-resolution.

Abstract: Animal re-identification (Re-ID) is a fine-grained recognition task that asks which gallery image depicts the same individual as a query image; one of its most immediate applications is the retrieval of lost companion animals from sightings databases. Existing methods report strong scores on curated benchmarks, but typically under a single seed, a single mask granularity, and without query-time corruption analysis, leaving open whether the reported gains survive realistic deployment conditions. We propose a hierarchical framework that decouples discriminative localisation (a YOLOv8 soft-crop) from identity embedding (a dual-stream network fusing a frozen CLIP ViT-B/16, adapted by a lightweight learned projection, with a fine-tuned ViT-Base carrying L2-norm part attention, trained jointly under ArcFace with PK sampling and a single BNNeck embedding). On a combined cat+dog open-set benchmark of 173 test identities, the proposed model attains Rank-1 = 0.9834 / mAP = 0.8731 at seed 42 and 0.9742 ± 0.0083 / 0.8597 ± 0.0185 over three seeds, surpassing a ViT-only ablation by+2.39 Rank-1 and +2.07 mAP percentage points and reducing the Rank-1 seed variance two-fold, although the mAP variance is not reduced. An open-set verification protocol shows that CLIP leads at the moderate operating points (AUC, EER, TAR@FAR= 0.01) but that all configurations converge near 68% TAR at the strictest FAR= 0.001. A background-bias evaluation brackets the residual background reliance of the embedding between a coarse bounding-box mask (∆R₁ = −1.65, a reliable lower bound) and a pixel-level SAM silhouette mask (∆R₁ = −19.9, ∆mAP = −28.8, an upper bound); a manual audit shows the silhouette segmenter destroys identity-relevant content in a non-trivial fraction of individuals, so the reliable soft crop is retained as the deployment front-end. A nine-corruption robustness audit identifies aggressive down-sampling (−16.0 Rank-1) and motion blur (−13.3) as the dominant single-source vulnerabilities. On the PetFace dog test split the same architecture reaches Rank-1 = 0.5414 over 14,716 unseen identities, and a genuine few-shot retraining regime (≥ 2 images per identity, K = 2) reaches Rank-1 = 0.5885 despite a 3.77× larger and sparser training pool. A Descriptor Vector Exchange (DVE) extension is ablated under full backbone backpropagation and is shown to be Pareto-dominated by the base model, a negative result we trace to the coarse feature-map resolution of the ViT backbone and to architectural redundancy rather than to the data.

Abstract: Processing tomato harvesters require real-time sorting of ripe tomatoes, unripe tomatoes, defective tomatoes, and soil clods in dense material flows. The key challenge is that category evidence is subtle, local, and spatially uneven: defect textures, fruit boundaries, and soil-clod cues are easily diluted by adjacent objects and background clutter. To address this, we propose SG-RTDETR, which adapts RT-DETRv2 to selectively allocate representation capacity to informative target regions while retaining local evidence. The model preserves high-resolution details during downsampling, applies saliency-guided token selection to focus global encoding on target-related locations, and reintegrates encoded tokens through residual spatial refill to maintain feature continuity. Context-aware feature organization and spatially adaptive cross-scale fusion further reduce fragmented responses and background-dominated fusion. Together, these operations target two main error sources: lost local cues and cluttered cross-region interactions. We constructed a four-category processing tomato dataset with 1,000 multi-object images and 400 single-object images for supplementary representation learning. Under a unified evaluation protocol, SG-RTDETR achieved 87.9% mAP50:95, 92.5% mAP50, and 95.6% mAR50:95, outperforming RT-DETRv2 by 3.4, 3.6, and 0.6 percentage points, respectively, while maintaining 103.6 FPS. These results show that selective modeling of spatially uneven evidence improves real-time tomato sorting.

Abstract: 3D printed textile nets are fabric-like structures produced by printing a few layers of continuous, overlapping paths. These textiles can be fabricated with minimal post-processing and material waste on low-cost Fused Deposition Modeling (FDM) 3D printers. Parametric design tools for generating toolpaths can be used to create a variety of textile structures and path patterns; however, the ability to control color patterns and transitions is limited by the continuity of the printed paths. We propose using halftoning to incorporate graphics patterns into the printed nets by dynamically controlling the printed path thickness during printing. Our workflow takes a grayscale input image and translates it into a two-color halftone pattern and a segmented toolpath. Line thickness is controlled by varying extrusion amount and printing speed, resulting in a colored pattern that matches the input image. We demonstrate our concept with a series of printed textile samples and a wearable corset.

Abstract: Robust object detection for autonomous driving requires a perception system that remains reliable when visible imagery is degraded by darkness, glare, rain, fog, motion blur, or long-range small targets. Visible and thermal infrared cameras provide complementary evidence, yet most existing RGB–thermal detectors still fuse modalities as aligned tensors and therefore underuse the relational structure hidden in channel responses, spatial layouts, semantic scales, and modality-specific uncertainty. This paper presents TopoGraph-Fusion, a hierarchical graph-guided dual-modal object detector that formulates fusion as topology-aware reasoning rather than feature concatenation. The proposed framework builds a dual-stream backbone for RGB and thermal images, constructs channel-wise topology through a Channel Topology Graph Aggregation module, derives relation-aware spatial and channel global attention from affinity graphs, and replaces fixed feature-pyramid communication with a Graph-Guided Feature Pyramid Network. A topology-regularized detection objective further encourages stable cross-modal correspondence while suppressing noisy all-to-all connections. Extensive draft experiments are organized on M3FD, FLIR, RGBTDronePerson, and VEDAI512, covering road scenes, adverse illumination, drone-person perception, and aerial vehicle detection. The results and visual analyses indicate that topology-guided fusion improves small-object recall, cross-modal consistency, and robustness under modality imbalance.

Abstract: Context and objectives: Lung ultrasound (LUS) is a safe low-cost tool that enables diagnosis, monitoring and guidance for interventional procedures at the patient's bedside. However, its expansion is hindered by a lack of training programs and the inherent difficulty of interpreting ultrasound images. In this context, Artificial Intelligence (AI) is emerging as a supportive tool for LUS interpretation, ensuring diagnostic efficacy and mitigating the shortage of experts. This systematic review aims to summarize and analyse recent advances in AI-based tools to support LUS interpretation. Methods: A systematic literature search was conducted across Web of Science, IEEE Xplore, and PubMed databases to identify peer-reviewed original journal articles published between 2015 and November 2025 that employed AI for the identification and localization of lung artifacts, anatomical structures, and pathological findings. Results: Twenty-four studies were included, identifying three main strategies: segmentation (10 studies), object detection (4 studies), and the generation of visual explanations through saliency maps (10 studies). All employed CNN-based architectures. The evaluation metrics used were heterogeneous. Conclusions: The development of AI systems to support LUS interpretation shows high potential; however, current studies exhibit significant heterogeneity in their objectives, methodologies, and evaluation metrics. It is necessary to move towards solutions designed for specific clinical environments and to adopt standardized protocols and evaluations that facilitate their implementation in clinical practice.

Abstract: Conventional Video Large Language Models (Video-LLMs) operate under an offline paradigm: they ingest a complete video before producing any response. This post-hoc design fundamentally limits their ability to serve as real-time interactors that perceive, reason, and communicate while the video is still unfolding. Streaming video understanding breaks this constraint by requiring models to process frames as they arrive and generate temporally grounded responses on the fly, thereby transforming passive narrators into active, always-on interactors. In this survey, we present a comprehensive and structured overview of recent advances in streaming video understanding, with a particular focus on Video-LLM-based approaches. We categorize existing methods into two complementary paradigms: proactive streaming models, which autonomously determine when to initiate responses in dynamic visual environments, and reactive streaming models, which focus on how to maintain long-term context and deliver efficient inference upon user queries over unbounded video streams. Together, these two paradigms characterize the core capabilities required for online video interactors: temporal awareness for timely engagement and scalable memory for sustained reasoning. We further provide a systematic review of emerging benchmarks and datasets designed to evaluate streaming-specific capabilities, including multi-turn dialogue, real-time captioning, and proactive response timing. Finally, we identify the key open challenges and outline promising future directions toward building truly interactive, efficient, and scalable streaming video systems. An accompanying resource repository is maintained at https://github.com/sotayang/Awesome-Streaming-Video-Understanding.

Abstract: This paper presents NexFire Pro, a deployable desktop framework for real-time fire and smoke surveillance across up to sixteen simultaneous IP or USB camera streams. The research contribution is a failure-mode-driven pipeline design methodology: each of five post-detection stages is derived from a distinct, identifiable failure mode category (illumination drift, static false-positive objects, spatially expected heat sources, temporal noise), and a camera-level hold-out protocol isolates each stage’s contribution independently, preventing the temporal data leakage that frame-level splits introduce. A fine-tuned YOLOv8n backbone serves as the fixed detection substrate, surrounded by a five-stage processing pipeline: (1) CLAHE and adaptive gamma correction; (2) MOG2-based motion validation; (3) neural detection; (4) normalised-coordinate ROI exclusion; and (5) temporal alarm hysteresis. Training images (14,872) were partitioned at the camera level into 70/15/15% splits to prevent temporal leakage. A systematic ablation study shows that the complete pipeline reduces the event-level false-alarm rate from 18.3% to 6.1% — a 66.7% relative reduction — against a YOLOv8n baseline on the same eight-camera held-out test set. Night-mode fire-detection precision improves by 19.8 percentage points on a 1,247-frame held-out near-infrared subset. Under eight concurrent 1080p streams the system sustains 11–13 analysed FPS at a 90th-percentile inference latency of 38 ms; under sixteen streams throughput falls to 5–6 FPS at 110 ms. CPU-only operation is characterised separately. Parameter sensitivity analysis confirms stability across ±20% perturbations of every key threshold. Beyond detector accuracy, the study also makes a methodological contribution: each pipeline stage is tied to a specific operational failure mode, and a camera-level hold-out protocol keeps frames from the same camera out of both training and test, which avoids the temporal leakage that frame-level splits introduce. Ablation experiments then measure how much each stage actually contributes to reliability, rather than reporting end-to-end numbers alone. The proposed design-and-evaluation framework is intended to be transferable to other reliability-oriented computer vision applications beyond fire detection.

Abstract: Deep neural networks achieve remarkable performance when training and test data share the same distribution, but this assumption often fails in real-world scenarios where data experiences continual distributional shifts. Continual Test-Time Adaptation (CTTA) tackles this challenge by adapting pretrained models to non-stationary target distributions on-the-fly, without access to source data or labeled targets, while addressing two critical failure modes: catastrophic forgetting of source knowledge and error accumulation from noisy pseudo-labels over time. In this comprehensive survey, we formally define the CTTA problem, analyze the diverse continual domain shift patterns that arise under different evaluation protocols, and propose a hierarchical taxonomy categorizing existing methods into three families: optimization-based strategies (entropy minimization, pseudo-labeling, parameter restoration), parameter-efficient methods (normalization layer adaptation, adaptive parameter selection), and architecture-based approaches (teacher-student frameworks, adapters, visual prompting, masked modeling). We systematically review representative methods in each category and provide comparative benchmarks and experimental results across standard evaluation settings. Finally, we discuss the limitations of current approaches and highlight emerging research directions, including adaptation of foundation models and black-box systems, offering a roadmap for future work in robust continual test-time adaptation, with further resources available at https://github.com/sarthaxxxxx/Awesome-Continual-Test-Time-Adaptation.

Abstract: The okra crop (Abelmoschus esculentus) plays a strategic role in family farming and traditional communities of the Brazilian semiarid region, but its productivity is strongly limited by root-knot nematodes (Meloidogyne spp.). The conventional phytosanitary diagnosis relies on manual egg counting under a Peters chamber, a slow, exhausting, and subjective process prone to inter-operator variability. This work presents PhytoNemaCount, a system that combines a low-cost, 3D-printed adjustable smartphone-to-microscope adapter with a Convolutional Neural Network (YOLOv8 Nano) to automate the detection and counting of Meloidogyne spp. eggs in microscopic images. The adapter, fabricated in Polylactic Acid (PLA) with an M5 threaded-rod fine-adjustment mechanism, enabled the standardized acquisition of 100 high-resolution images (960×1280 px) from guava (Psidium guajava) root samples naturally infected with Meloidogyne spp., collected in São José da Tapera, Alagoas, Brazil. A controlled stability experiment demonstrated that the adapter reduced inter-frame centroid displacement tenfold relative to free-hand smartphone capture (3.1 ± 0.8 px vs. 31.4 ± 9.2 px). Data augmentation expanded the effective training set to approximately 350 instances per epoch. Images were annotated on the MakeSense.ai platform; the YOLOv8n model was trained for 50 epochs under a 70/15/15 (train/val/test) split, achieving a precision of 0.988, a recall of 0.875, an F1-score of 0.928, and a mAP50 of 0.967 on the held-out test set. A pilot reproducibility study showed that the automated system achieved a coefficient of variation (CV) of 4.2% across repeated counts of the same slide set, compared with a mean CV of 18.7% observed among three independent human operators, confirming substantially superior counting reproducibility. Two user interfaces were implemented: a Streamlit web application for batch processing of static images and an interactive Tkinter/DroidCam module for real-time detection directly from the microscope. These results confirm the technical feasibility of converting conventional optical microscopes into low-cost standardized digital capture stations for computer-vision-based phytosanitary diagnostics, in alignment with Agriculture 4.0 principles.

Abstract: Zooplankton serve as critical bioindicators of marine ecosystem health, yet their accurate classification remains challenging due to subtle inter-class differences, substantial intra-class variations, and complex underwater imaging conditions. While Vision Transformers (ViTs) have shown promise in fine-grained recognition, they struggle with the local feature modeling and multi-scale perception essential for zooplankton identification. This paper proposes ViT-MDFA, a novel architecture that synergistically integrates Multi-scale Dilated Convolution (MSDC) and Dual Attention (DA) mechanisms into the Vision Transformer framework. The MSDC module employs parallel dilated convolutions with strategically selected dilation rates to capture both fine-grained textures and global structural patterns without computational overhead. The DA mechanism combines channel-wise and spatial attention to adaptively emphasize diagnostically relevant features while suppressing background interference. Extensive evaluations across four challenging benchmarks (WHOI-Plankton, ZooScanNet, Kaggle-Plankton, and Dec-22) demonstrate that ViT-MDFA achieves state-of-the-art performance, reaching 97.46\% accuracy and 96.73\% F1-score on the Dec-22 dataset—surpassing the baseline ViT-B/16 by remarkable margins of 6.14\% and 7.79\%, respectively. Comprehensive ablation studies validate the individual contributions of MSDC (+1.83\%) and DA (+0.81\%) modules, with sensitivity analysis identifying the optimal dilation rate configuration [6,12,18]. Grad-CAM visualizations reveal that ViT-MDFA consistently attends to biologically meaningful morphological structures, such as antennae, body segments, and caudal spines. The proposed architecture achieves this superior performance while maintaining a lightweight, modular design suitable for deployment on flow cytometers and edge computing platforms, thereby enabling real-time, automated zooplankton monitoring for marine ecological assessment.

Abstract: Huizhou Intangible Cultural Heritages, represented by wood carving, brick carving, stone carving, and paper cutting, have intricate patterns and stylistic diversity. Existing image classification approaches rely heavily on manual curation, lacking scalability and interpretability for large-scale digital exhibitions. This study proposes a deep learning framework that adapts large-scale vision–language models to Huizhou Intangible Cultural Heritages through information-maximization self-distillation. Specifically, we employ Contrastive Language-Image Pre-Training (CLIP) as a teacher model to generate pseudo-labels for unlabeled Huizhou heritage images, while a student model is adapted through parameter-efficient tuning of visual prompts and LayerNorm parameters. An entropy-based information maximization objective further enhances model confidence and diversity. Extensive experiments demonstrate improvements over zero-shot and existing adaptation baselines. Overall, MM-SHAP offers a novel, information-theoretic paradigm for unsupervised and efficient fine-tuning in exhibition-focused visual classification tasks, balancing performance and efficiency.

Abstract: Industrial anomaly detection, the computer-vision core of automated quality inspection, had by 2023 settled into a stable method taxonomy organized around a single assumption: a separate detector trained per product on defect-free images alone, evaluated on closed-set benchmarks. The 2024-2026 wave breaks this assumption. Web-pretrained foundation models (vision encoders, multimodal large language models, and diffusion models) generalize across product categories without per-product retraining, recasting anomaly detection from per-category density estimation into a foundation-model-prior problem: detection from frozen pretrained features, defect reasoning in natural language, mixture-of-experts models shared across modalities, and language-controlled synthesis of the missing abnormal data. Because the reconstruction, embedding, distillation, and memory taxonomy of earlier surveys no longer absorbs this literature, we reorganize it around one question (what replaces category-specific normal-only training?) into five emerging generalization mechanisms: visual, reasoning, geometric and multimodal, universal, and synthesis priors. We then audit the field's benchmark and evaluation frontier, catalog eight cross-cutting bottlenecks that no current method resolves at once, and propose a five-problem research agenda. We conclude that, despite real progress on all five mechanisms, current methods still generalize within curated benchmarks more than under real-world deployment; closing that gap is the agenda's central aim.

Abstract: Pedestrian and face detection in unconstrained surveillance scenarios faces significant challenges due to low-resolution targets, complex backgrounds, and varying illumination conditions. This paper proposes an improved Multi-task Cascaded Convolutional Neural Network (MTCNN) framework for simultaneous pedestrian and face detection in unconstrained surveillance scenarios, with particular emphasis on low-resolution target detection. The proposed method introduces three key innovations: (1) a restructured “pedestrian-first, face-second” cascade architecture that exploits the spatial correlation between pedestrian bodies and faces to reduce search space; (2) a multi-level feature fusion mechanism in the face detection subnet that integrates high-level semantic features with low-level spatial details to enhance detection of tiny and blurry faces; and (3) a pyramid pooling network with RoIPooling layers in the final stage, enabling scale-invariant processing of multi-scale targets with Batch Normalization for improved classification stability. Extensive experiments demonstrate that the proposed algorithm achieves a pedestrian detection rate of 34.2% on small-scale targets (Far subset, Caltech dataset) at 38 FPS, outperforming YOLO (22.8%) and Faster R-CNN (28.5%), and a face detection rate of 63.1% on the WIDER Face Hard subset, representing a 2.4 percentage-point improvement over the original MTCNN (60.7%). The unified detection framework achieves an overall processing speed of 31 FPS, satisfying real-time requirements for surveillance applications.

Abstract: Automated corrosion inspection is essential in the maritime industry, where continuous monitoring of ship hulls and docking infrastructure is necessary to ensure operational safety and structural reliability. However, many existing predictive and deep learning-based approaches depend heavily on large annotated datasets, extensive training procedures, and substantial computational resources, limiting their practicality in real-world and resource-constrained inspection environments. To address these challenges, this work presents a compact and fully unsupervised corrosion segmentation framework for ship docking images based on graph-based similarity modeling and community-aware clustering. The proposed approach constructs pixel-level structural representations using lightweight graph descriptors, performs coarse clustering to obtain initial segmentation labels, and refines spatial consistency through Leiden community detection. Unlike conventional unsupervised segmentation techniques that frequently over-smooth fine structures or generate overly simplified partitions, the proposed method preserves localized corrosion patterns while effectively suppressing segmentation noise. We compare the framework against established unsupervised segmentation methods, including graph-based segmentation, probabilistic aggregation, mean-shift segmentation, and normalized cuts, using compactness, silhouette coefficient, and edge density as evaluation metrics. Experimental observations demonstrate that the proposed approach achieves improved structural delineation and boundary preservation while remaining entirely training-free and computationally efficient. By leveraging graph-based similarity matching without reliance on supervised learning, the framework offers a scalable and interpretable solution for practical industrial inspection scenarios. The lightweight design further supports potential adaptation to broader structural monitoring applications, including pipelines, offshore systems, and large-scale infrastructure inspection.

Abstract: Hepatic and splenic flexures are important anatomical landmarks in colonoscopy, supporting segmental orientation, lesion localisation, and bowel preparation assessment. This exploratory study investigated whether liver- and spleen-related shadow patterns visible on the colonic wall could serve as indirect cues for flexure localisation. We developed a preliminary workflow combining colour-based annotation and deep learning semantic segmentation. Representative shadow and mucosal regions were used to extract colour characteristics, while lumen/borders and artefacts were identified using luminance-based criteria. Based on these rules, 500 colonoscopy frames with visible shadows were automatically annotated and used to train a segmentation network for four classes: shadows, intestinal mucosa, lumen/borders, and artefacts. The trained model was then applied frame by frame to colonoscopy videos to generate temporal profiles of the segmented regions. In a representative test video, shadow clusters appeared near expert-indicated flexure regions after empirical filtering of frames not matching several criteria. These preliminary findings suggest that liver- and spleen-related shadows may provide complementary cues for flexure localisation in selected cases. However, the method remains exploratory and requires validation on larger, independently annotated datasets.

Abstract: Ovarian cancer remains a leading cause of gynaecological cancer mortality, and ultrasound based deep learning systems for its diagnosis are typically built as separate post-hoc processes for classification, segmentation, and interpretability, which introduces work flow inefficiencies and may produce inconsistent predictions. This work addresses that limitation. We propose UM-TOTA (Unified Multi-Task Ovarian Tumor Architecture), a Vision Transformer (ViT) based architecture that performs eight-class tumor classification, three-class malignancy detection, tumor segmentation, and clinical concept interpretability within a single unified framework. We integrate a concept bottleneck guided by the IOTA and O-RADS clinical guidelines to enable transparent decision-making through medical concepts that clinicians can understand, and we employ combined adaptive t-vMF Dice 11 and boundary-enhanced segmentation losses with progressive task weighting to stabilise multi-task optimisation. We evaluated the model on the Multi-Modality Ovarian Tumor Ultrasound (MMOTU) 2D dataset using 5-fold stratified cross-validation. UM-TOTA achieved 80.26% ± 1.10% accuracy (97.06% one-vs-rest macro specificity) for eight-class classification, 90.88% ± 1.14% accuracy (90.41% specificity) for malignancy detection, and 77.29% ± 1.29% Dice for segmentation, while reducing the computational parameter load by approximately 66.7% relative to sequential single-task pipelines. The learned concepts aligned with established malignancy criteria, identifying vascularization, solid components, and papillary projections as key predictors. This unified approach offers an efficient and interpretable framework for clinical ovarian ultrasound workflows.