Visible-thermal tiny pedestrian detection in UAV aerial images is crucial for online decision-making in urban security and disaster response. However, the extremely small scale and sparse distribution of pedestrians cause discriminative cues to be submerged by dominant low-frequency background and contextual redundancy during feature learning. Meanwhile, cross-modal spatial misalignment and spatially varying modality reliability hinder stable fine-grained correspondence, thereby degrading fusion quality. To address these issues, QA2FDet is proposed as a quality-aware adaptive alignment fusion network comprising three modules: spectrum-spatial decoupled enhancement module (SDE), cross-modal correspondence mining module (CCM), and prior-informed gated fusion (PGF). SDE leverages the discrete cosine transform to disentangle redundant low-frequency background information, while deep semantic gating propagates high signal-to-noise ratio details into shallow representations to enhance subtle cues of tiny pedestrians and suppress high-frequency noise. To establish fine-grained neighborhood correspondences under slight spatial offsets, thermal-guided local asymmetric cross-attention is designed in CCM. Finally, region-level quality and modality discrepancy are jointly modeled for adaptive cross-modal fusion in PGF. Extensive experiments on multiple UAV-based RGBT detection benchmarks demonstrate that QA2FDet achieves state-of-the-art performance and exhibits strong robustness in challenging aerial scenes.
Insects inhabit complex vibroscapes shaped by substrate-borne vibrations from multiple biotic and abiotic sources. One underappreciated topic is how vibrations function in predator-prey interactions. Tiny warty birch caterpillars (Falcaria bilineata) are known to produce complex vibratory signals to defend leaf-tip territories against conspecifics, raising the question of whether vibratory signalling and sensing also play roles in predator-prey interactions. We staged encounters between resident neonate caterpillars and three natural intruders: conspecifics, ladybird beetle larvae and adult ladybird beetles, while simultaneously recording behaviour and substrate-borne vibrations. Resident caterpillars showed three key responses - vibratory signalling, freezing and dropping - but these responses varied strongly with intruder identity and stage of encounter. Residents signalled vigorously toward conspecifics, with rates escalating as intruders approached their territories. In contrast, encounters with predators evoked predator-specific defensive strategies including freezing and dropping. Adult ladybird beetles, which caused high mortality (43%), elicited rapid escape responses and subsequent territory abandonment, while ladybird beetle larvae, which caused no mortality, triggered slower responses with initial signalling that ceased upon closer approach. Critically, vibrations generated by the 'footsteps' of each approaching intruder type produced distinct vibratory 'signatures', differing in amplitude, spectral and temporal characteristics. Resident caterpillars also initiated defensive responses before physical contact, often when intruders were still centimetres away. Together, these findings demonstrate that these miniature larvae, no larger than ∼1-2 mm, thrive in complex vibroscapes where vibrations not only function to advertise territory ownership against conspecifics but also provide essential early-warning cues enabling sophisticated threat assessment and context-appropriate defensive responses in predator-rich environments.
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
With improving survival, periviable neonates (≤25 weeks' gestation) represent a dynamic, but under-studied population in neonatal care, with persistently high cardiopulmonary and vascular vulnerability. Immature cardiovascular structure and function as a consequence of immature myocardial architecture, altered calcium handling, relative adrenal insufficiency, and persistent fetal shunts contribute to complex and dynamic cardiovascular physiology in this population. This may be present clinically in the form of hypotension and low end-organ perfusion. Traditional paradigms of blood pressure-based definitions of hypotension are poorly validated in this population and do not accurately reflect systemic blood flow or end-organ perfusion. Emerging evidence supports a phenotype-based multi-parametric approach to cardiovascular assessment, to distinguish ongoing physiological changes such as ductal physiology, pulmonary hypertension, low systemic vascular resistance and primary myocardial dysfunction phenotypes. However, significant knowledge gaps remain, including lack of normative hemodynamic values, and limited evidence guiding pharmacologic therapies. This narrative review focuses on the cardiovascular challenges in the management of periviable neonates as they transition to extrauterine life, delineating cardiac phenotypes, describing modalities of cardiovascular assessment and identifying existing knowledge gaps. We propose a physiology-based approach to cardiovascular management strategies based on existing, albeit limited, evidence. IMPACT: Periviable neonates present unique hemodynamic challenges due to structural and functional cardiovascular system immaturity, which can be categorized into different hemodynamic phenotypes dictated by baseline cardiac function, lung compliance and directionality of intracardiac shunts, especially the patent ductus arteriosus. In the absence of established normative reference values for common modalities of cardiac assessment, optimal care should consist of early identification of cardiac phenotypes, continuous surveillance, physiology-based management strategies, and frequent reassessment to guide individualized treatment.
Understanding human evolution relies on biomolecular data from ancient skeletal tissues, yet warm climates often cause complete collagen loss, excluding many regions from the study. This research investigates the survival of non-collagenous proteins (NCPs) and low-molecular-weight proteins in archaeological bones deemed collagen-free by traditional metrics. Using a multi-method approach, we employed sandwich enzyme-linked immunosorbent assays for osteocalcin quantification, Qubit fluorometry for total protein, and liquid chromatography tandem mass spectrometry (LC-MS/MS) for characterization, complemented by Fourier transform infrared spectroscopy to assess the diagenetic state. Samples included collagen-depleted bones from Neolithic Lebanon and Palaeolithic France, with well-preserved controls from Neolithic Serbia and Paleolithic Russia. Results indicate that bone-associated NCPs, including osteocalcin, survive only if the insoluble collagen is preserved. Methodologically, tangential flow filtration outperformed centrifugal devices for protein recovery. EDTA demineralization with FASP was most effective for maximizing collagen recovery, while HCl demineralized protein precipitation best detected unique NCPs. Collagen was identified in the soluble supernatants of most collagen-depleted bones. Abundant collagen peptides were identified in a sample with a 0% collagen yield and a very low amide-to-phosphate ratio. These findings demonstrate that bones unsuitable for traditional dating can still retain measurable collagen, broadening the range of biomolecular analyses possible in warm and humid climates.
Microbial biodiversity is essential for the proper functioning and balance in the ecosystem, and they possess numerous biotechnological and industrial applications. Microbial biobanks serve as a major resource for conserving valuable microbial diversity, providing access for education, research, and industrial applications. The present comprehensive review article discusses the history, responsibilities, and applications of microbial biobanks, covering different regulatory ecosystem and their guidelines for the governance of biobanks. It specifically provides the current global status of MCCs and precisely sheds light on MCCs in India. This review critically highlights a stark disparity in the bio-resources management, showing dominance of Asia with substantial investment in bio-infrastructure. In India, a significant geographical disparity was observed with a heavy concentration of culture collections in states with established research infrastructure. Despite the immense commercial potential of microbes, very few MCCs are held by the private sector globally, leaving the burden of conservation on the public sector and raising concerns about long-term survival and financial instability. Furthermore, there is a critical need for emerging advanced preservation strategies, AI-assisted tools for biobanking operations, and stringent quality control procedures in biobanks meeting international standards. In addition, the findings underscore the necessity of the digitalization of culture collections to facilitate global data sharing and robust policies for access and benefit sharing of microbial resources. In conclusion, addressing these regional disparities, mitigating financial and infrastructure gaps, using emerging and advanced strategies, and digitalization of CCs are imperative for the sustainability and applicability of the microbial biobanks.
暂无摘要（点击查看详情）
暂无摘要（点击查看详情）
Phage infection undergoes a series of physiological transitions, holding crucial information about phage replication dynamics and potential phage-derived antimicrobials. Although phage-induced cytological changes have been used to infer phage-hijacking mechanisms, current approaches are limited by the lack of comprehensive single-cell morphological analysis and insufficient resolution of temporal dynamics, particularly for phages displaying short latent periods, thereby hindering systematic characterization of morphological transitions throughout the infection cycle. Here, we characterized a newly identified coliphage with a genome of 53 kbp, Tiny, which exhibits an unusually long latent period, making it an ideal candidate for resolving temporal morphological transitions. Tiny exhibits both temperature- and host-dependent killing profiles against diverse Escherichia coli strains, including ATCC 25922, uropathogenic E. coli, and avian pathogenic E. coli. While its replication efficiency is temperature-dependent, showing enhanced productivity at lower temperatures, the duration of its adsorption and latent period is largely host-dependent. Depending on the bacterial strain, Tiny exhibits either a prolonged latent period in slow-adsorbing strains or a rapid one in fast-adsorbing strains, regardless of infection temperature, suggesting phage-host compatibility. Single-cell bacterial cytological profiling of Tiny-infected ATCC 25922 cells revealed that Tiny progressively induces distinct bacterial morphological transitions throughout its lytic cycle, suggesting sequential interference with host physiology as part of its replication cycle prior to cell lysis. This work establishes a broadly applicable framework for dissecting lytic phage biology with high temporal resolution and lays the foundation for future integrative omics studies aimed at understanding how phages sequentially modulate their bacterial hosts. Antibiotics trigger unique patterns of morphological changes in bacteria, and these compound-specific signatures provide a basis for determining mechanisms of action in antibiotic discovery. By the same concept, phage-induced morphological changes can reveal key insights into phage replication dynamics and guide the identification of phage-derived antimicrobials. However, the complexity of phage biology and the variability of phage-host interactions pose challenges in interpreting these phenotypic outcomes. Here, we employed a phage-host pair that exhibits an unusually prolonged latent duration as a model to establish a broadly applicable framework for dissecting lytic phage biology with high temporal resolution. Through single-cell bacterial morphological analyses, this approach captures dynamic infection processes inducing morphological transitions across the phage replication cycle. This work provides a phenotypic analysis pipeline to advance our understanding of phage-host interactions and lays the foundation for future integrative omics studies to elucidate how phages sequentially modulate their bacterial hosts.
The domesticated apple (Malus domestica) is widely considered an admixture primarily involving M. sieversii, M. sylvestris, and M. orientalis, with possible contributions from tiny-fruited species such as M. baccata. Although this origin theory is well supported by genomic studies, the relative contributions of each progenitor species to cultivar genomes remain unclear, and taxonomic classifications of some Malus species are conflicting. The availability of new genomic tools and resources enables probing of these topics. A panel of wild Malus accessions was genotyped using the Illumina Infinium 20K apple SNP array to generate a library of species-specific SNP and haploblock alleles. After excluding heavily admixed individuals and masking admixture haplotypes in individuals with small amounts of admixture, this library was used to clarify the ancestry of 436 wild accessions and estimate the genomic proportions attributable to M. sieversii, M. sylvestris, M. orientalis, and collectively, tiny-fruited Malus species in M. domestica cultivars. The results were broadly consistent with previous studies but revealed new ancestry details and enabled precise haplotype-level attributions. Previously unreported admixture and taxonomic irregularities were detected for some wild accessions. All M. domestica cultivars were found to be mixtures of the three primary progenitors, with some also showing contributions from tiny-fruited Malus species. The relative contributions varied among cultivars, with evidence that older cultivars and traditional cultivars originating in eastern Europe or Western Asia had less M. sylvestris ancestry than modern cultivars and traditional cultivars originating in Western Europe. This study provides new insights into apple ancestry and highlights the need for clarification in Malus taxonomy. The findings have implications for germplasm management, historical research, and genetic improvement of apple cultivars. The species-specific allele library developed here offers a valuable resource for routine admixture estimation of Malus accessions genotyped on the same set of SNPs.
Background/Objectives: The aim of this study is to test different convolutional neural network (CNN) and Transformer-based models to detect and classify vertical misfit at the abutment-prosthesis interface on panoramic radiographs, and to develop a hybrid deep learning model enhanced with attention mechanisms. Methods: A dataset consisting of a total of 566 images, manually classified as 249 'fit' and 317 'misfit' cases by two experts, was created. Images were resized to 224 × 224 and divided into training, validation, and test groups. The deep learning model yielding the most successful results was determined as the backbone; a hybrid model was developed by integrating three different attention modules (SE, CBAM, and ECA) into this structure. Model performance was evaluated using accuracy, precision, sensitivity, and F1 score metrics. Results: CNN-based models (RegNetY-800MF, ConvNeXt-Tiny, EfficientNetV2-S, ResNet50) performed better than Transformer-based models (DeiT, Swin-Tiny) in all metrics. The proposed hybrid model exhibited the highest success among all tested models with a 99.12% accuracy rate. This model reached a 100% precision value in the misfit group and yielded no false positive results. The F1 scores of the hybrid model were recorded as 99.01% for the fit group and 99.21% for the misfit group. Conclusions: The findings of this study demonstrate that attention-enhancing deep learning frameworks have the potential to significantly improve the diagnostic utility of routine panoramic radiographs. It shows that panoramic imaging, when supported by advanced artificial intelligence, can provide valuable diagnostic support in detecting vertical misfit. The developed model has the potential to become a reliable clinical decision support system.
Oral squamous cell carcinoma (OSCC) is the most common malignancy of the oral cavity, and early diagnosis plays a crucial role in improving patient prognosis and survival rates. Histopathological examination remains the gold standard for OSCC diagnosis; however, this process is time-consuming and highly dependent on expert interpretation. With the rapid development of digital pathology and artificial intelligence, deep learning-based approaches have emerged as promising tools to support automated diagnostic systems. In this study, five convolutional neural network (CNN) architectures-VGG16, ResNet50, InceptionV3, EfficientNetV2S, and ConvNeXt-Tiny-were comparatively evaluated for the automated classification of OSCC using histopathological images. An open-access OSCC dataset was utilized, and two experimental scenarios were created using the original dataset and an augmented dataset. The dataset was divided into training, validation, and test subsets using a stratified approach. All models were trained under identical experimental conditions using ImageNet-pretrained weights and a unified classifier head in order to ensure a fair comparison. Model performance was assessed using Accuracy, Precision, Recall, Specificity, F1-Score, and ROC-AUC metrics. Additionally, Grad-CAM was applied to visualize the image regions influencing model predictions and to enhance interpretability. The results demonstrated that data augmentation significantly improved the performance of all models. Among the evaluated architectures, ResNet50 achieved the highest diagnostic performance on the augmented dataset, reaching an accuracy of 0.91 and a ROC-AUC of 0.88, followed by EfficientNetV2S and ConvNeXt-Tiny. Visualization analyses indicated that the models focused on histopathologically relevant regions associated with tumoral structures. Overall, the findings suggest that deep learning-based approaches can effectively support patch-level automated OSCC classification from histopathological images and may contribute to future development of clinical decision support systems in digital pathology.
In this paper, we tackle a core challenge for wearable human activity recognition (HAR), namely the recognition of daily-living and locomotion activities from inertial sensor windows: delivering reliable, interpretable uncertainty on tiny microcontrollers where latency, RAM, and energy are tightly constrained. Existing embedded approaches either calibrate softmax confidences, which are cheap but brittle under sensor placement or tempo shifts, or rely on generative or posterior-sampling schemes that exceed TinyML budgets. We propose hyperdimensional distance- and uncertainty-aware human activity recognition (HDUQ-HAR), an on-device hyperdimensional computing (HDC) framework that encodes each IMU window into a bipolar hypervector, classifies via prototype similarity, and derives three complementary, lightweight uncertainty signals from the same representation: (1) distance to the class prototype, (2) similarity gap to the runner-up, and (3) vote-dispersion capturing n-gram consensus. A label-conditional conformal layer converts these scores into set-valued predictions with finite-sample coverage guarantees and exposes a human-readable reason code indicating why uncertainty increased. Across UCI HAR, WISDM, PAMAP2, and OPPORTUNITY with subject-out splits and realistic shifts (orientation, gain, time-warp, missing axis, cross-placement), HDUQ-HAR achieves near-target coverage at [Formula: see text] with near-singleton sets on i.i.d. data (average size 1.18-1.25) and robust shift/OOD detection (AUROC 0.92-0.96), while running in 3-5 ms/window on Cortex-M4 with ∼6-9 KB RAM and ∼5-7 KB Flash. By unifying HDC geometry with label-conditional conformal prediction, our method shows that efficiency and reliability can co-exist in wearables, yielding small, calibrated sets that expand gracefully under shift and actionable explanations practitioners can trust.
Clear cell renal cell carcinoma (ccRCC) is a highly aggressive malignancy with a high rate of recurrence and limited therapeutic options. Carbonic anhydrase IX (CA9) is characteristically overexpressed on the surface of ccRCC cells, making it a promising target for site-specific drug delivery. However, identifying the key molecular drivers of ccRCC progression and developing efficient, targeted nanomedicines remain critical challenges in current research. Bioinformatics analysis of TCGA and single-cell RNA sequencing data was used to elucidate the ETS1/MYC axis. Direct transcriptional regulation of MYC by ETS1 was experimentally validated by chromatin immunoprecipitation-quantitative PCR (ChIP-PCR) and dual-luciferase reporter assays. An optimized CA9-targeting peptide, CaIX-P7, was designed via computational modeling and mutation screening, with affinity validated by surface plasmon resonance (SPR). siETS1-loaded liposomal nanoparticles (LNPs) were prepared using microfluidics and surface-functionalized with CaIX-P7 (ETS1@Lip-CAIX). The nanoparticles were characterized for size, zeta potential, and encapsulation efficiency. Therapeutic efficacy was evaluated in ccRCC cell lines (786-O, A-498), patient-derived organoids (PDO), and nude mouse xenograft models. Single-cell analysis identified ETS1 and MYC as synergistically activated transcription factors within tumor epithelial cells. Mechanistically, ChIP-PCR and dual-luciferase assays demonstrated that ETS1 promotes MYC transcription through this specific binding site, establishing ETS1 as a direct transcriptional activator of MYC. The optimized peptide CaIX-P7 demonstrated superior binding affinity to CA9 (Kd=52.96 nM) compared to its precursor. The engineered ETS1@Lip-CAIX nanoparticles exhibited a stable size of 154.8 nm and high siRNA encapsulation efficiency (89.1%). Systematic evaluation revealed that ETS1@Lip-CAIX effectively silenced the ETS1/MYC axis, leading to significant growth inhibition across all models, including patient-derived 3D organoids and in vivo xenografts, showed no discernible morphological alterations or pathological damage in major organs. This study identifies the ETS1/MYC axis as a novel therapeutic target in ccRCC. We further successfully developed a CA9-targeted nanoplatform, ETS1@Lip-CAIX, which exhibits robust anti-tumor efficacy by disrupting this newly discovered regulatory hub. These findings provide a foundation for future translational studies of ccRCC. Clear cell renal cell carcinoma (ccRCC) is the most aggressive and common type of kidney cancer. Despite improvements in surgery and existing drugs, many patients face high recurrence rates and limited treatment options. Through advanced data analysis and laboratory experiments, we identified a critical growth engine in kidney cancer cells called the ETS1/MYC axis. This pathway acts as a switch that drives rapid tumor cell multiplication. Our study found that patients with high levels of these factors typically have a poorer prognosis. To shut down this growth engine, we engineered a smart delivery system called ETS1@Lip-CAIX. This system consists of tiny lipid-based carriers (nanoparticles) loaded with a genetic drug (siRNA) designed to silence the ETS1 gene. To ensure the drug reaches its target, we attached a specialized GPS peptide (CaIX-P7) to the surface. This peptide specifically recognizes and binds to CA9, a protein that is overexpressed on the surface of kidney cancer cells but rarely found in healthy kidneys. We tested this nanomedicine using diverse models, including traditional cell cultures, patient-derived 3D organoids (mini-tumors that mimic real patient biology), and animal models. Results demonstrated that the targeted nanoparticles significantly inhibited tumor growth by disrupting the ETS1/MYC axis. Crucially, the treatment showed no toxicity to normal cells, indicating high safety and biocompatibility. This research identifies a novel therapeutic target and presents a precise, highly effective nanomedicine strategy. It provides a promising new avenue for personalized treatment of aggressive kidney cancer, potentially improving survival rates while minimizing side effects for patients.
Accurate medical image segmentation requires both fine boundary localization and robust contextual understanding, which is often difficult to achieve simultaneously, particularly in lightweight architectures. In this paper, we propose SwiftMSeg, a lightweight encoder-decoder framework that integrates a convolutional encoder, a transformer-based local-global-local module, and a hierarchical multi-scale decoder. The proposed framework addresses the boundary-context challenge by effectively combining progressive multi-scale refinement for fine boundary separation with global context modeling through long-range dependency aggregation. Extensive evaluations on publicly available colonoscopy, pathology, ultrasound, and magnetic resonance imaging datasets demonstrated the capability of SwiftMSeg to accurately segment diverse anatomical structures, ranging from tiny nuclei to polyps and large tumor regions. The model further demonstrated moderate domain-independent generalization on an external dataset, achieving Dice scores of 0.896 (colonoscopy), 0.860 (pathology), 0.850 (ultrasound), and 0.870 (MRI), consistently outperforming most baseline methods. In addition, it achieved improved boundary localization with lower Hausdorff distance (e.g., 16.43 in MRI and 33.89 in ultrasound) and reduced average symmetric surface distance, indicating more precise and stable segmentation. Statistical analysis further confirmed that the improvements of SwiftMSeg are significant ([Formula: see text]) with large effect sizes across modalities, validated by both paired t-tests and Wilcoxon tests. Despite its strong performance, SwiftMSeg remains highly efficient, requiring only 4.48M parameters and 0.940 giga floating-point operations per second (GFLOPs), reducing computational cost by approximately ∼53× compared to the U-Net-based baselines (standard U-Net ∼31M parameters and ∼50 GFLOPs), while maintaining high segmentation accuracy. These results highlight the effectiveness of SwiftMSeg as a practical and scalable solution for real-world medical image segmentation across diverse modalities.
This article presents a comprehensive analysis of a tiny, rectangular, and exponentially tapered slotted ultra-wideband (UWB) Vivaldi antenna developed explicitly for brain microwave imaging. The antenna comprises seven symmetric rectangular slots carved on both sides of the rear radiative wings of a tapered slot antenna. The antenna's total dimensions, which encompass the ground layer, are 0.39λ × 0.33λ × 0.007λ while operating at a lower frequency of 1.35 GHz. The simulated impedance bandwidth (IBW) is observed to have a range of frequencies between 1.35 and 5.50 GHz, with a fractional bandwidth (FBW) of 121.17%, where the magnitude of |S11|< - 10 dB. On the other hand, the measured IBW spans from 1.4 to 6.4 GHz, exhibiting an FBW of 128.20%. The antenna has a gain of 9.99 decibels isotropic (dBi) and possesses directional radiation properties, along with a front-to-back ratio (FBR) exceeding 30 dB. The analysis conducted in the time domain demonstrates little pulse distortion and a group delay of less than one nanosecond, indicating good performance. Additionally, the system exhibits a high-fidelity factor of 96.26% and maintains linearity in the transmission phase (S21). Both simulation and experimental findings demonstrate the efficacy of the antenna for microwave-based imaging.
Snakebite envenoming is a significant global health crisis that has been long neglected as a global health priority. It is a huge problem for rural communities of low and middle-income countries, India accounts for the largest proportion of snakebite deaths globally. Timely identification of venomous snakebite and its syndromic pattern is essential for effective administration of antivenom and supportive treatment. Expert identification of snake species and syndromes is not always available in peripheral healthcare settings. This leads to delays, unnecessary referrals, or improper treatment choices. Additionally, diverse snake species distribution and venom variations across regions pose challenges. AI-powered image classification methods can help overcome these barriers. We propose a clinically oriented deep learning pipeline for binary classification of venomous and non-venomous snake species of India using real-world imagery data. This pipeline would serve as a baseline step towards aiding snakebite management at peripheral healthcare setups with scarce resources. The selected dataset consisted of 20 medically important Indian species. MobileViT-S, ConvNeXt-Tiny, EfficientNet-V2-S and ResNeXt-50 (32 × 4d) were trained under same conditions for comparison of results. Model interpretability was evaluated using Grad-CAM ++ to ensure that classification was not performed based on background but on features like head shape and stripes present on body. For reliable implementation we connected it to a web interface with human in loop expert verification. Experts can confirm or override predictions in real time. Among the evaluated architectures, ResNeXt-50 (32 × 4d) showed the most reliable and consistent performance in classifying venomous and non-venomous snakes. It achieved the highest test accuracy, sensitivity, specificity, and F1-score. The model also had strong discriminative ability, with a ROC-AUC of 0.9950 and PR-AUC of 0.9959. These results indicate dependable performance in safety-critical screening situations. Grad-CAM++ visualizations confirmed that predictions were based on anatomically relevant features, especially in the head and body contour areas. This supports model interpretability and reduces background bias. Although the dataset size and single-institution source limit how widely the results can be applied, the proposed framework shows that it's possible to create a clinically oriented, ready-to-use deep learning system for snakebite triage support. This system is intended as a scalable tool to help rural healthcare workers, emergency responders, and telemedicine platforms in areas where snakebites are common.
Deep learning for mammographic image classification yields impressive performance metrics, but inconsistent evaluation methodologies-specifically whether results are reported at the independent side level or the bilateral patient level-make cross-study comparisons unreliable. The aim of this study was to quantify, on a single dataset and with a uniform training recipe, how much of the reported performance is determined by evaluation granularity rather than by model architecture. We benchmarked six backbone architectures (ResNet-18, ResNet-50, EfficientNet-B0, DenseNet-121, ConvNeXt-Tiny, ViT-B/16) crossed with three multi-view fusion strategies (concatenation, bilateral asymmetry, cross-view spatial attention) on the biopsy-confirmed Chinese Mammography Database (CMMD; 706 four-view patients), using five-fold patient-level stratified cross-validation. Sixteen configurations completed training for both binary malignancy diagnosis and five-class BI-RADS assessment. We report side-level and patient-level metrics; statistical analyses include 5-fold Wilcoxon signed-rank tests, DeLong's paired AUC test on pooled per-case scores, and bootstrap 95% confidence intervals. Side-level AUC exceeded patient-level AUC by an average of 17.5 percentage points (range 12.7-22.4), an effect that dwarfs the absolute differences observed between CNN backbones (<3 AUC points). DeLong tests resolved approximately half of all CNN-vs-CNN pairwise comparisons at p<0.05 despite small effect sizes, whereas ViT-B/16 underperformed every CNN variant by 8-10% AUC despite having 6-10× more parameters. Patient-level multi-class BI-RADS evaluation under the standard probability-averaging aggregation rule returned a degenerate macro-AUC of exactly 0.000-a property of the metric/aggregation pair, not of the models-and three concrete alternative aggregators are proposed. The extreme patient-level malignancy prevalence intrinsic to this biopsy-confirmed cohort (96.2%) rendered all models unable to identify non-malignant patients at the patient level. Reporting methodology, evaluation granularity, and dataset composition are compounding confounds in mammography classification research. Absolute performance numbers reported on CMMD should not be extrapolated to population screening settings, where prevalence is several orders of magnitude lower; studies should report both side-level and patient-level metrics with mutually consistent label/aggregation rules, and characterise performance using confidence intervals or paired statistical tests rather than fold-level Wilcoxon alone. On the CMMD four-view subset, side-level AUC exceeds patient-level AUC by an average of 17.5 percentage points (range 12.7–22.4), an effect that dwarfs the absolute differences observed between CNN backbones (<3 points).The standard patient-level multi-class BI-RADS aggregation rule (label =max, score = mean of probability vectors) returns a degenerate macro-AUC of 0.000 under bilateral asymmetry; this is an artifact of the metric/aggregation pair, not a measurement of model capability. Three concrete alternative aggregators are proposed.Among CNN backbones, pooled-score DeLong tests resolve approximately half of all pairwise comparisons at p<0.05, but absolute AUC differences are small (≤0.030); the 5-fold Wilcoxon test cannot resolve any CNN-vs-CNN pair due to its p=0.0625 floor.ViT-B/16 underperforms all CNN variants by 8–10 % AUC and fails on a partially different patient population (Jaccard 0.29 vs. 0.42 for CNN-CNN pairs), consistent with the data-hungry behavior of vision transformers on small medical datasets.CMMD is a biopsy-confirmed enrichment cohort (96.2 % malignant at the patient level); absolute performance numbers should not be extrapolated to population screening, where prevalence is several orders of magnitude lower.Per-case predictions for all 16 (backbone, fusion) configurations on both tasks are released alongside the manuscript, enabling further methodological work without retraining.