Search Results (242)

Rail-surface defects are often small, low-contrast, and confused with steel texture or reflections, making it difficult to improve accuracy without increasing model cost. This study proposes ESiV-YOLOv8, a compact detector that assigns complementary modifications to feature selection, box regression, and multiscale fusion in YOLOv8n. EffectiveSE recalibrates deep backbone features, SIoU provides direction-aware regression, and a VoV-GSCSP/GSConv neck reduces redundant computation. Evaluation used a self-built four-class dataset of 4020 images, a near-duplicate-aware train-validation-test split, and matched seven-seed experiments. ESiV-YOLOv8 achieved 97.7% precision, 94.6% recall, 97.2% mAP@0.5, and 68.6% mAP@0.5:0.95. Relative to YOLOv8n, mAP@0.5:0.95 increased by 5.0 percentage points, while parameters and GFLOPs decreased by 17.1% and 12.2%, respectively. On the combined natural-condition subset, ESiV-YOLOv8 achieved 59.6% mAP@0.5:0.95, 5.6 percentage points above the baseline. The annotation audit yielded 98.0% class agreement and a mean box IoU of 0.89. Model-only and end-to-end latency increased by 1.4% and 2.2%, respectively. Overall, ESiV-YOLOv8 improves detection accuracy and reduces model scale with limited latency overhead on the tested backend. Full article

This study investigates how sentiment polarity relates to offensive function in Chinese online attacks, focusing on cases where offence is expressed implicitly rather than through explicit negative wording. Using a 1400-item Chinese online-text corpus derived from TOXICN and its pragmatics-oriented annotation extension, the study examines sentiment distributions across non-offensive texts, explicit attacks, and implicit attacks, and evaluates how these patterns are reflected in model-generated sentiment labels. The implicit-attack subset is analyzed through four pragmatic strategies: Irony, Trope, Indirectness, and Exaggeration. A stratified 400-item validation subset was independently annotated by three human annotators for sentiment polarity, and four NLP models were used to assign sentiment labels to the full corpus: Claude Haiku 4.5, OpenAI GPT-4.1-mini, a Chinese BERT-based sentiment classifier, and Multilingual DistilBERT. Human reference annotations showed that negative sentiment was strongly associated with offensive-language labels but did not map onto them perfectly: some non-offensive texts were perceived as negative, while some implicit attacks were assigned neutral sentiment. Pragmatic strategy did not significantly predict human reference sentiment polarity in the validation subset. Model–human agreement varied across the four selected systems, with Claude Haiku 4.5 and OpenAI GPT-4.1-mini showing higher agreement than Chinese BERT and Multilingual DistilBERT in the validation subset. Full-corpus mixed-effects logistic regression showed that model-generated negative labels varied significantly by text category and model, whereas evidence for stable model-specific pragmatic-strategy effects was limited. These findings provide empirical evidence from Chinese online attacks that sentiment polarity can inform, but cannot replace, offensive-language analysis. They also show that model-generated sentiment labels require human validation and cautious interpretation when offence is implicit, figurative, homophonic, or context-dependent. Full article

Asian elephants (Elephas maximus) face substantial thermoregulatory constraints because of their large body size, low relative surface area, sparse hair, and lack of functional sweat glands. Reliable body temperature measurement is essential for assessing thermal status and evaluating welfare in both wild and managed populations, but conventional rectal thermometry requires close physical contact, animal training, and repeated manual handling, making high-frequency, continuous, large-scale monitoring impractical. This study developed a non-invasive framework for automatically detecting the outer ear and extracting auricular surface temperature from infrared thermograms. Rectal temperature, regional surface temperatures, ambient temperature, and relative humidity were measured synchronously in eight semi-captive Asian elephants, yielding 425 matched observations. The associations between rectal temperature and the surface temperatures of three anatomical regions (head, outer ear, torso and limbs) were analyzed using repeated-measures correlation accounting for the non-independence of repeated measurements. Mean outer-ear temperature showed the strongest within-individual association with rectal temperature (r_rm = 0.395, p < 0.001), identifying the outer ear as the optimal thermal window for subsequent automated monitoring. Eight lightweight YOLO models—YOLOv5n, YOLOv5s, YOLOv8n, YOLOv8s, YOLO11n, YOLO11s, YOLO26n, and YOLO26s—were trained on 2178 annotated infrared images and evaluated on an independent 194-image test set from extra elephants. Model performance was assessed using detection metrics, inference speed, Bland–Altman agreement, Taylor diagram statistics, and a weighted multi-criteria score with Monte Carlo sensitivity analysis. YOLO11n achieved the best overall performance, with an mAP50 of 0.933 and an inference speed of 164 frames per second. The proposed framework provides an efficient method for automated auricular temperature monitoring and has potential applications in elephant welfare management and remote physiological surveillance. Full article

HERA-GM separates an offline learned tactical proposal from the authority to execute it. A Dueling Double DQN trained with Conservative Q-Learning proposes one of five tactical actions. The runtime process then uses the action-value margin, semantic feature availability, annotation density, a Mahalanobis diagnostic, and fixed hard-rule conditions to assign ACCEPT, DEFER, or RECOVER. The study separately examined proposer agreement, authority changes, closed-loop outcomes, and held-out-family discrimination. The original frozen evaluation used 113 nuPlan Mini scenarios from 39 logs and 1970 closed-loop runs. An additional exploratory behavior-cloning block added 339 runs. Behavior cloning had slightly higher offline macro-F1 than CQL, whereas DDQN without CQL had much lower agreement under the tested configurations. The main comparison between M1 and the simpler B3 gate showed no supported primary safety difference, indicating limited added endpoint effect from Mahalanobis and hard-rule evidence in this cohort. M1 also showed lower safety-failure and drivable-area violation rates than behavior cloning, but with lower conditional progress; the primary result did not remain below 0.05 after pooled Holm adjustment across the five clean comparisons. The Mahalanobis score did not distinguish held-out semantic families reliably. The findings describe the operating trade-offs and limits of conditional execution authority rather than a safety guarantee. Full article

Reliable abnormal behavior recognition from surveillance videos is hindered by the high cost of clip-level annotation, the scarcity of abnormal samples, and the context-dependent nature of behavioral semantics. Although vision–language models offer strong semantic transferability, their application under limited supervision remains susceptible to noisy pseudo-labels and confirmation bias. We propose confidence-aware semi-supervised vision–language contrastive learning (CA-VLC), which jointly exploits limited labeled videos and abundant unlabeled videos. Building on an existing CLIP-initialized temporal backbone, CA-VLC combines behavior-only and context-enriched text prototypes through confidence- and agreement-guided semantic fusion. For unlabeled videos, the model generates predictions from weakly augmented views and selects reliable pseudo-labels using entropy-based confidence estimation and class-adaptive thresholds. Detached weak-view targets then supervise strongly augmented views through confidence-weighted self-training without requiring an additional teacher network. Furthermore, cross-view consistency regularization and confidence-aware contextual alignment suppress unreliable semantic cues and improve robustness to contextual noise. Experiments on CABR50 demonstrate consistent improvements across multiple labeled-data ratios, while evaluations on CABRZ6 and UCF-101 assess prompt-based transfer to predefined target label sets without target-domain fine-tuning. With 10% labeled videos, CA-VLC achieves 84.06% Top-1 accuracy and 83.51% Macro-F1, retaining 95.47% of its fully supervised Top-1 accuracy of 88.05%, thereby demonstrating its effectiveness for label-efficient abnormal behavior recognition. Full article

Background: Recent pan-cancer transcriptome analyses have identified ENST00000453316 as a cancer biomarker candidate, showing consistent and significant deregulation across multiple cancer types. Methods: This study investigated the transcript as a regulatory RNA in cancer-associated processes and its potential as a biomarker using a bioinformatics approach. Results: In silico characterization classified ENST00000453316 as a non-coding transcript with regulatory features and stable structure. Curated annotations documented the binding of 21 RNA-binding proteins across the transcript region, with core spliceosomal factors at the unique splice junction and cytoplasmic 3′-end factors at the terminal exon, in agreement with the sequence-based predictions. Promoter analysis revealed a TATA-less, CpG-poor region with multiple putative transcriptional regulator binding sites. Its expression was higher in most solid tumor types analyzed, with minimal expression in non-tumor tissues, indicating a tumor-associated pattern. Discrimination between tumor and adjacent non-tumor tissue within TCGA was good in five of the ten tissue types with sufficient adjacent-normal samples, with the area under the ROC curve reaching 0.82 in the lung, 0.78 in the bladder and 0.71–0.75 in the colon, breast and stomach, but close to chance in the kidney, thyroid and prostate. Conclusions: These results suggest that ENST00000453316 is a cancer-associated, promoter-driven non-coding transcript with potential regulatory relevance. Despite its revised annotation, it exemplifies how alternative promoter usage generates biologically and clinically meaningful transcripts contributing to tumor-specific gene regulation. Future experimental validation is needed to confirm the role of ENST00000453316 as a functional regulatory RNA and assess its clinical potential as a cancer biomarker. Full article

Background/Objectives: To develop and evaluate an automated CT-based framework for the quantitative assessment of fibrotic interstitial lung disease (ILD), including idiopathic pulmonary fibrosis (IPF), using a standardised six-level anatomical protocol and deep-learning lung segmentation. Methods: The segmentation dataset comprised 3315 manually annotated development slices from 92 patients and a non-overlapping internal holdout of 845 slices from 5 patients. A separate 100-study localisation/scoring set yielded a 97-patient agreement cohort (84 IPF, 13 other ILD; 1164 per-level, per-lung observations) after three DICOM-conversion exclusions. YOLO11n-seg masks underwent vessel- and structure-removal fibrosis detection. The radial spatial score was compared with a non-blind expert-adjudicated reference; the per-level Fibrosis Index was an auxiliary read-out. Results: On the five-patient internal segmentation holdout, mean intersection over union (mIoU) was 0.926 ± 0.017; the in-sample development value was approximately 0.95. Model-only latency was 73.8 ± 9.7 ms/slice at batch size 1, and peak throughput was 1.48 ms/slice at batch size 512. In the separate 97-patient agreement cohort, the expert-adjudicated score was identical to the automated score for 967 of 1164 observations (83.1%) and differed for 197 (16.9%). In the modified-score subset, Pearson r was 0.918, mean absolute error was 3.07, and ICC(2,1) was 0.889 (patient-clustered 95% CI 0.828–0.923). The pooled ICC(2,1) was 0.988 (0.982–0.992), but this value was inflated because the 967 unchanged pairs were identical by construction. Sequential end-to-end processing, measured in seven study patients, took a mean of 22.5 s per patient (median 24.0 s, range 17.3–24.9 s); localisation accounted for 88.1% of this time. Conclusions: The framework combined lung segmentation, anatomically standardised sampling, and automated fibrosis scoring. The radial score showed preliminary analytical concordance under non-blind expert adjudication. The fibrosis detector remains a proof-of-concept implementation based on 8-bit windowed images and has not been compared with independently drawn pixel-level fibrosis masks. The radial partition is an exploratory scoring convention and was not compared with alternative partitions or validated against clinical outcomes. Larger external studies using native Hounsfield-unit data and independent blinded readers are required. Full article

Accurate and scalable driver performance assessment is critical for improving road safety and reducing traffic-related injuries and fatalities, particularly in low- and middle-income countries where the majority of global road deaths occur. This paper presents an exploratory proof-of-concept framework for automated driver evaluation that combines real-world dashcam footage, YOLOv8-based object detection, and multimodal large language models (MLLMs), specifically Gemini 1.5 Flash. Two prompting strategies, narrative and rule-based, were designed to assess driver behavior against standardized licensing criteria derived from the California Department of Motor Vehicles (DMV) driving performance evaluation score sheet. The framework was evaluated across 11 manually curated driving scenarios covering intersections, pedestrian crossings, stop signs, cyclists, and emergency vehicles. Ground-truth labels were established through consensus between two traffic engineering experts cross-referencing official California DMV evaluation criteria. In this preliminary evaluation, the rule-based prompt achieved higher agreement with ground-truth assessments (10/11 scenarios, 90.9%) compared to the narrative prompt (7/11 scenarios, 63.6%), particularly in detecting clear rule violations. The narrative approach demonstrated greater contextual flexibility in ambiguous situations. These results should be interpreted as preliminary, given the small sample size, manually curated dataset, and absence of large-scale statistical validation. Nonetheless, the findings illustrate how combining visual detection with structured language-model prompting may support interpretable, policy-aligned driver evaluation. Key limitations include dependence on video quality, limited scenario diversity, absence of temporal behavioral modeling, and reproducibility constraints tied to proprietary API behavior. Future work should expand validation to larger annotated datasets, incorporate temporal sequence modeling, and explore region-specific regulatory adaptation. Full article

The growing use of artificial intelligence (AI)-based language technologies in learning, assessment, writing support, and speech-enabled educational platforms has made linguistic representation an increasingly important dimension of sustainable and inclusive education. This exploratory mixed-methods study examines how transcribed forms of Turkish dialects are processed within task-appropriate natural language processing (NLP) environments and how their digital representation is interpreted by linguists and dialect speakers. The study combines a spoken-language corpus compiled from 100 speakers in 14 provinces across Türkiye’s seven geographical regions with spaCy- and Stanza-based morphosyntactic outputs, an exploratory BERTurk-based semantic-similarity component, and semi-structured interviews. The corpus contains approximately 60 h of recordings, 120 pages of transcripts, and 3500 annotated structures. Descriptive task-specific observations indicate recurrent processing difficulties involving compound tense forms, non-canonical word order, dialect lexicon, discourse particles, and pragmatically marked expressions. For the POS configurations examined, descriptive POS agreement was higher on standard written Turkish (SWT) reference material (89–91%) than on transcribed dialect data (60–65%). The qualitative material indicates that non-recognition may be interpreted as cultural invisibility, reduced trust in digital tools, and possible pedagogical misalignment. These educational and social dimensions are presented as potential implications rather than directly tested outcomes. The study conceptualizes digital linguistic sustainability through three interrelated dimensions: technical robustness in processing Turkish dialects, the continuity of culturally embedded linguistic knowledge, and educational inclusion. Full article

Background: Sleep loss is epidemiologically linked to atherosclerotic cardiovascular disease, yet shared-transcriptome studies typically intersect differentially expressed gene lists and treat the overlap as a shared response, assuming rather than testing directional agreement. Methods: Controlled human sleep restriction (GSE39445; 438 whole-blood samples from 26 participants) and paired carotid plaque versus patient-matched intact artery (GSE43292; 32 patients) were compared across biological compartments at three resolutions: individual genes, annotated gene sets, and redundancy-collapsed programs. The resulting architecture was transferred unmodified to acute total sleep deprivation, plaque progression, and aneurysm enlargement, and was evaluated by participant-grouped cross-validation, network topology, exact-cis Mendelian randomization with colocalization, and single-cell localization. Results: Convergence was resolution dependent. The cohorts shared more differentially expressed genes than expected under independence (274 of 16,869; p = 3.54 × 10⁻⁴), yet only 48.9% changed concordantly (p = 0.66). Concordance reached 78.0% among jointly significant gene sets and 70.0% (35 of 50) after redundancy collapse (bootstrap 95% CI, 0.560–0.820), coupling neutrophil-granule and myeloid-effector enrichment with depletion of RNA-processing, chromatin, and ciliary programs. Under a contrast-independent maximum interquartile range probe-selection rule, the gene-level overlap was smaller and no longer nominally significant (190 genes; p = 0.082), whereas the prespecified 50-program architecture, rescored without re-clustering, retained the same 35 concordant programs. The architecture generalized to acute sleep deprivation (91.4%) and plaque progression (97.1%) but inverted during aneurysm enlargement (20.0%). Prespecified separability was borderline (AUC 0.558; 95% CI, 0.486–0.630; permutation p = 0.052). Of twelve topology-prioritized spliceosomal candidates, SF3A3 alone combined FDR-significant exact-cis Mendelian randomization (OR 0.851; 95% CI, 0.772–0.939) with regional colocalization (PP.H4 = 0.831). Conclusions: Sleep loss and atherosclerosis converge as coordinated transcriptional programs rather than as shared individual genes, although the magnitude and statistical enrichment of gene-level overlap were sensitive to representative probe selection; this represents cross-compartment transcriptional convergence rather than replication and does not confer individual-level separability. SF3A3 is an experimentally testable candidate, not a validated therapeutic target. Full article

Robotic bin-picking of disordered, randomly stacked workpieces remains challenging because reliable grasping depends on an accurate estimate of object pose, yet many established solutions require high-precision 3D sensing, detailed object models, or large annotated datasets that raise the cost and effort of deployment on a new production line. This work presents a complete binocular vision framework that estimates workpiece pose by image matching and executes vision-guided grasping on a 6-DOF manipulator. A pose-annotated multi-view template library is constructed automatically through robot-driven image acquisition and compressed by a coarse-to-fine clustering scheme, and object pose is estimated by discriminative template matching with rigid refinement. To characterize the geometric reliability of the matched poses, an offline cross-modal analysis relates the 2D templates to a 3D reference model of the object and measures their agreement through region and contour reprojection metrics. Grasp configurations are then generated under orientation and collision constraints and corrected online by closed-loop visual feedback. Experiments on two representative workpieces show template-matching accuracy of 89–90% against classical and learned similarity measures, and grasp success between 81 and 87% across single-object and mixed scenes, outperforming the GraspNet baseline under the tested conditions. The framework offers an accurate and deployment-friendly route to robotic bin-picking. Full article

This work presents a reproducible YOLO-based pipeline for bottle detection in sandy environments, emphasizing dataset integrity, leakage-free evaluation, and deployment-oriented model selection. A one-class dataset of 1585 images and 3167 annotated bottles was audited to identify annotation-format defects and near-duplicate contamination between training and validation partitions. Sequence membership was reconstructed through perceptual-image similarity and used to assign complete image components to a sequence-aware train/validation split, eliminating the near-duplicate pairs found in the initial random partition. A controlled ablation holding model, seed, and corrected labels fixed showed that the random split reports 0.040 higher mAP@0.5:0.95 than the sequence-aware split (0.787 vs. 0.747), quantifying the leakage risk directly rather than only asserting it. Five YOLO configurations were then benchmarked under three independent seeds each; the observed mAP@0.5:0.95 differences among models (0.004–0.008) were small in absolute magnitude and, given only three seeds per model, are interpreted descriptively rather than as evidence of statistical equivalence or significance, so yolo11n_bottle was selected through a joint accuracy-parity, compactness, and exportability criterion (precision 0.982, recall 0.985, mAP@0.5 0.992, mAP@0.5:0.95 0.748), using approximately ten times fewer parameters than the largest configuration and producing a 5.2 MB checkpoint. ONNX export preserved detection geometry closely (100% count agreement, mean matched IoU $\ge 0.9998$), without meeting strict metric-parity tolerances. A stratified sample of 108 frames from operational RealSense BAG footage was manually annotated by an independent reviewer and evaluated quantitatively: mAP@0.5 remained close to the internal validation figure (0.927 vs. 0.992), while mAP@0.5:0.95 fell substantially (0.483 vs. 0.747), revealing a localization gap between the curated benchmark and operational conditions that this manuscript reports transparently. Together, these results show that dataset auditing, sequence-aware partitioning, multiseed benchmarking, and manually annotated operational evidence are each necessary to interpret a detection benchmark built from continuous video acquisition, providing a traceable, reproducible workflow for selecting and evaluating compact visual-perception models for resource-constrained environmental applications. Full article

Accurate and efficient counting of Oudemansiella raphanipes plays a key role in intelligent cultivation management, yield estimation, and unmanned production monitoring. However, automated detection of Oudemansiella raphanipes remains challenging due to dense fruiting body distribution, morphological variations, complex soil backgrounds, and illumination fluctuations in practical cultivation environments. To address these challenges, this study proposes OR-YOLO, an enhanced deep learning-based detection framework for in situ fruiting body recognition and counting. The proposed model integrates Bi-level Routing Attention (BRA) and Coordinate Attention (CA) modules to improve multi-scale feature representation and spatial localization capability, while an FEIoU-VFL optimization strategy is introduced to enhance bounding-box regression and confidence estimation for difficult samples. Experimental results demonstrated that OR-YOLO achieved Precision, Recall, mAP50, mAP75, mAP50–95, and F1-score values of 88.2%, 84.5%, 89.5%, 72.2%, 65.8%, and 86.3%, respectively. Compared with the baseline model, OR-YOLO effectively reduced missed detections caused by dense growth and target occlusion. Furthermore, the model maintained stable performance under different soil backgrounds and illumination conditions. The predicted fruiting body counts showed strong agreement with manual annotations, with a coefficient of determination (R²) of 0.945 and a correlation coefficient (r) of 0.972, demonstrating the reliability of OR-YOLO for automated quantitative phenotypic analysis. Collectively, this study provides an efficient and robust solution for intelligent mushroom monitoring and offers a detection approach for automated phenotyping in complex agricultural production environments. Full article

Automated extraction of structured threat information from unstructured cyber threat intelligence (CTI) underpins modern security operations, yet the supporting machine learning resources are almost exclusively English: no annotated Arabic CTI named entity recognition (NER) corpus has been published. We introduce AraCTI-NER, a dataset of 10,312 token-level annotated samples (275,530 tokens; 42,360 entity spans) over eight STIX-inspired entity types, built by an LLM-assisted pipeline seeded with authentic Arabic cybersecurity articles, structurally validated and rebalanced through targeted generation. We benchmark seven encoders from three families (Arabic-specialized, English cybersecurity-adapted, and multilingual) over three seeds under strict entity-level metrics, and release a 408-sentence expert-audited test subset (ATS-gold) whose reliability is quantified by a second independent expert validation (inter-annotator agreement 0.878 entity-level F1). XLM-RoBERTa Large attains the best mean F1 (0.7603; 0.7674 on ATS-gold), with AraBERTv2 close behind (0.7491), while both English-only cybersecurity encoders fall to ≈0.63, a separation that holds across every seed and survives expert correction, with the ≈3-point F1 decrease from ATS-silver to ATS-gold concentrated in Vulnerability and TTP. On 350 doubly annotated sentences from authentic Arabic cyber-incident news, a shift in both provenance and register, the strongest model reaches F1 = 0.5429 against an inter-annotator F1 of 0.616. AraCTI-NER establishes the first reproducible baseline for Arabic CTI NER and identifies domain-adaptive Arabic cybersecurity pre-training as the highest-value next step. Full article

Background/Objectives: Tissue fold artifacts observed in hematoxylin-eosin (H&E)- stained preparations used in histology education can complicate the assessment of normal tissue architecture and affect students’ accurate interpretation of microscopic structures. This study aimed to automatically detect and localize tissue fold artifacts in histology teaching preparations using deep learning methods. Methods: A total of 2127 hematoxylin and eosin (H&E)-stained histological images of brain, kidney, liver, small intestine, and testis tissues obtained at 10× magnification were used in the study. The dataset consisted of 899 clean/artifact-free images and 1228 images containing tissue fold artifacts. Seven deep learning architectures, including convolutional neural networks (CNN)-based models and a vision transformer-based model, were evaluated for image-level classification: ResNet18, ResNet50, DenseNet121, EfficientNet-B0, EfficientNet-B3, ConvNeXt-Tiny, and Swin-Tiny. Classification performance was evaluated using an organ-based testing approach, a slide-level train–test split in which images from the same histological slide group were retained within a single subset, a general image-level 80/20 train–test split, and five-fold cross-validation. A DeepLabV3-ResNet50-based segmentation model was trained using QuPath-prepared masks to determine fold regions at the pixel level. Grad-CAM was used for qualitative visualization and quantitative comparison with manually annotated fold regions. Results: All classification models showed excellent performance overall. The Swin-Tiny model was the most accurate in terms of general image-level classification, scoring 99.06% for accuracy, 99.19% for F1-score, and 99.98% for AUC. In slide-level classification, EfficientNet-B3 showed the best performance in terms of accuracy (99.54%) and AUC (99.99%). No pairwise statistical difference was found among the models using Holm adjustment. In the five-fold cross-validation setting, ResNet50 achieved the best performance with the accuracy of 98.73 ± 0.54%. In the detailed small-intestine error analysis, classification accuracy was 67.67%, with high sensitivity (98.59%) but low specificity (24.34%), mainly because of false-positive predictions. In the segmentation evaluation, the final DeepLabV3-ResNet50 model trained with combined BCE + Dice loss resulted in Dice of 0.7630 ± 0.2425 and IoU of 0.6661 ± 0.2577 on the independent test set. False positive segmentations were rare in 899 artifact-free images (only 0.33% of images had a tissue fold region of at least 1%). The quantitative Grad-CAM analysis showed poor spatial agreement with the manually annotated tissue fold masks (Dice = 0.2423; IoU = 0.1441). In the independent MPP10 dataset, ResNet50 achieved 89.63% accuracy and 88.49% F1-score. Conclusions: The suggested method was able to detect tissue fold artifacts as well as localize them in the teaching images of histology. The strong performance recorded at the slide level validates the internal findings, but the limited performance in the case of small intestine images and on the external dataset reveals that tissue structure remains a crucial factor. Full article