Search Results (782)

Objectives: Preterm infants confront multiple coexistent health risks and may not receive human milk (HM), which contains human milk oligosaccharides (HMOs), among other biologically active, beneficial components. 2′-Fucosyllactose (2′-FL) is the most abundant HMO in most HM. Interventional studies of formula with added HMOs have demonstrated multiple beneficial outcomes in healthy, term infants. Methods: This double-blind, randomized, placebo-controlled pilot study (NCT03306316) evaluated the effect of a 2′-FL supplement on growth, tolerance, and neurobehavior, as assessed by the Neonatal Intensive Care Unit (NICU) Network Neurobehavioral Scale (NNNS), in preterm infants. Eligible neonates born between 26 0/7 and 31 6/7 weeks gestational age (GA) were enrolled from three NICUs in the United States. Preterm infants consuming mother’s own milk or donor human milk were randomized to receive a supplement of 0.135 g/mL of 2′-FL or 0.05 g/mL dextrose (placebo) twice daily until 36 weeks corrected GA, NNNS assessment, or hospital discharge, whichever came first. Results: Of 42 participating preterm infants, 23 (55%) received 2′-FL and demonstrated no significant differences in growth in weight, length, or head circumference over 5 weeks. There were also no differences in stool frequency, adverse (AEs) or serious adverse events (SAEs), or neurobehavior (NNNS) between groups. Conclusions: This pilot study found that 2′-FL supplementation in preterm infants through 36 weeks corrected GA was safe and well-tolerated and supported adequate growth and neurobehavior. Full article

Background: Osteoradionecrosis of the jaw (ORNJ) remains a major late complication of head and neck radiotherapy, and risk assessment relies mainly on clinical and dosimetric factors. We constructed the Host Tissue Repair Vulnerability Index (HTRVI), a composite biomarker integrating inflammatory, immune-nutritional, and oxygenation-related parameters, to estimate biological susceptibility to ORNJ in locally advanced nasopharyngeal carcinoma (LA-NPC). Methods: This retrospective cohort included 261 LA-NPC patients treated with definitive concurrent chemoradiotherapy between 2010 and 2021. HTRVI was calculated as (CRP × platelet × neutrophil)/(albumin × lymphocyte × hemoglobin). HTRVI was compared with hemoglobin (Hb) and the Global Immune-Nutrition-Inflammation Index (GINI) using receiver operating characteristic analysis. Multivariable logistic regression, restricted cubic spline analysis (RCS), and bootstrap resampling assessed independent association, continuous risk relationship, and internal validation. Results: During a median follow-up of 63.8 months, 24 patients (9.2%) developed ORNJ. HTRVI showed superior discrimination (AUC, 0.864; 95% CI, 0.769–0.932) compared with Hb (AUC, 0.785; p = 0.001) and GINI (AUC, 0.759; p = 0.045). HTRVI remained independently associated with ORNJ after adjustment for mandibular mean dose and post-CCRT tooth extraction burden (OR per standard deviation increase, 4.81; 95% CI, 1.10–20.99; p = 0.037). RCS analysis showed a significant continuous association between HTRVI and ORNJ risk (P_overall < 0.001) without nonlinearity (P_nonlinear = 0.736). Internal validation showed minimal optimism (bootstrap-corrected AUC, 0.993). Conclusions: HTRVI was independently associated with ORNJ and provided additional predictive information beyond established clinical and dosimetric factors in this single-center cohort. The continuous HTRVI–ORNJ association supports a biological continuum model of host tissue repair vulnerability. However, HTRVI should currently be regarded as an investigational biomarker requiring independent external and prospective validation before clinical implementation. Full article

Background: Psychiatric inpatients may be vulnerable to head trauma; however, associations among psychiatric diagnosis, trauma mechanisms, and acute trauma-related neuroimaging findings among patients requiring neurosurgical consultation remain unclear. Methods: This retrospective observational study included psychiatric inpatients referred for neurosurgical consultation after head trauma at a tertiary neuropsychiatric specialty hospital between July 2022 and December 2025. Psychiatric diagnoses were grouped into psychotic, mood, substance use, organic/symptomatic, and other disorders. Trauma mechanisms were classified as falls, self-harm, assault/fight-related injuries, and other trauma types. Radiological findings were classified as trauma-related abnormalities or secondary incidental findings, including age-related changes. Associations were evaluated using chi-square tests, with false discovery rate correction applied to global and binary comparisons. Exploratory multivariable regression models were used as sensitivity analyses. Results: The cohort comprised 461 psychiatric inpatients who sustained head trauma, underwent radiological imaging, and were referred for neurosurgical consultation. Trauma mechanisms differed significantly across psychiatric diagnostic groups (χ² = 30.89, p = 0.002). Falls were the most common mechanism (189/461, 41.0%). Assault/fight-related trauma showed the clearest diagnosis-associated pattern (χ² = 19.72, p < 0.001, Cramér’s V = 0.207) and remained the only significant binary outcome after false discovery rate correction. Acute trauma-related imaging findings were observed in 35 patients (7.6%) and were not associated with psychiatric diagnosis (χ² = 6.16, p = 0.187). Calcifications were the most frequent imaging finding (191/461, 41.4%). Conclusions: In psychiatric inpatients, while head trauma patterns and neuroimaging findings have the potential to be explained by psychiatric diagnoses, they may also reflect unmeasured behavioral and clinical factors. Full article

To address the reliance on empirical knowledge and the lack of a quantitative basis for worker positioning in high-altitude UHV live-line work, a refined power plant model consisting of workers, transmission towers, insulator strings, and conductors has been established. Based on altitude-corrected breakdown field strength and discharge attachment criteria, entry paths and positioning for the cradle and power-driven ascender methods were optimized under IVI- and VVV-type insulator strings. Results show that field concentration is more pronounced in the middle phase. For the cradle method, a 30° approach angle is optimal, reducing body surface field strength by 9.4–18.6%, with discharge points at the toes; a sitting posture with slightly raised legs and toes pointing toward the conductor is recommended. For the power-driven ascender method, vertical entry from the side outperforms entry from below, reducing the far-tower side-phase field strength by 22–24%, with discharge points at the head; an upright, sideways posture with the face turned away from the conductor is recommended. This study provides a quantitative basis for high-altitude UHV live-line work. Full article

Background and Objectives: Advances in perinatal and neonatal care have improved survival in preterm infants, yet neurodevelopmental impairment remains a major concern. A considerable proportion of preterm infants with normal or only mildly abnormal neuroimaging findings still develop neurodevelopmental difficulties, highlighting the need for more sensitive imaging markers beyond the detection of major brain lesions. Cranial ultrasound is a widely used bedside imaging tool that enables both detection of brain injury and quantitative assessment of brain growth. This study aimed to evaluate the association between early and term-equivalent cranial ultrasound morphometric parameters, head circumference, and neurodevelopmental outcomes assessed by the Bayley Scales of Infant and Toddler Development, Fourth Edition (Bayley-4). Materials and Methods: This exploratory observational study included 22 preterm infants admitted to a neonatal intensive care unit. Cranial ultrasound examinations were performed within the first 72 h of life and at term-equivalent age (37–40 weeks postmenstrual age). Measured parameters included the Levene index, lateral ventricular width, third ventricle width, subarachnoid space, and interhemispheric fissure. Head circumference was analyzed as an additional marker of postnatal brain growth. Neurodevelopment was assessed at 12–14 months corrected age using Bayley-4. Associations were analyzed using Spearman’s rank correlation. Results: Significant correlations were more frequent at term-equivalent age than in the early neonatal period. Early measurements showed positive associations between ventricular parameters and motor outcomes, as well as between head circumference and receptive communication (p < 0.01). Third ventricle width was also associated with receptive communication. At term-equivalent age, lateral ventricular width demonstrated consistent negative correlations with motor and language outcomes, with strong associations in motor domains (p < 0.01). Head circumference showed limited positive associations, reaching significance only for personal–social skills, while third ventricle width was associated with play and leisure activities. Conclusions: Morphometric cranial ultrasound measurements, particularly those obtained at term-equivalent age, may provide valuable additional information about early neurodevelopment in preterm infants. Head circumference reflected postnatal brain growth but showed less consistent associations with neurodevelopmental outcomes than ventricular morphometric parameters. Larger prospective studies with longer follow-up are needed to confirm these findings and define the prognostic value of individual cranial ultrasound morphometric parameters and head circumference for predicting neurodevelopmental outcomes in preterm infants. Full article

Quantitative phylogenetics in historical linguistics has relied almost entirely on lexical cognate data. This study asks a different question: how much genealogical signal can be recovered from structural features extracted from annotated corpora, and whether it survives at deep time depths. We compute 29 structural features—including Shannon entropies of dependency direction and of dependency-relation distributions, relation-specific directionality ratios, dependency-distance measures, and constructional ratios—across 25 Transeurasian languages from the five proposed groups (Turkic, Mongolic, Tungusic, Japonic, and Koreanic) and three outgroups (Chinese, Vietnamese, and Hindi), 28 languages in all. Most of the Tungusic and Mongolic languages have no running-text corpus, so we built new Universal Dependencies treebanks for them by glossing example sentences from reference grammars; thirteen are used here. Each feature was tested for phylogenetic signal (Pagel’s λ and Blomberg’s K, with FDR correction) under four competing reference topologies, and the features that passed were used for tree inference (Bayesian inference in MrBayes, with Neighbor-Joining as a check). The same pipeline was first run on Indo-European in a companion study, where it recovers only individual subgroups and does not resolve a stable tree. At the depth proposed for the Transeurasian family (a Proto-Transeurasian root of about 9000 years before present), the structural signal was not enough to reconstruct the family’s internal relationships. The signal tests favoured a flat three-way division of the major branches (7 strict/20 relaxed features) over any nested hypothesis (≤2 strict features each), and the strongest signal lay in core word-order parameters (e.g., object direction, λ = 1.00, K = 6.06). But both Bayesian and distance-based inference returned near-complete polytomies: although the chains converged (ASDSF < 0.01), no branch reached a posterior probability above 0.75, and none of the three multi-language branches (Turkic, Mongolic, or Tungusic) was recovered. The outgroup test made the reason clear: Hindi, which is Indo-European but SOV, grouped with the head-final Transeurasian languages rather than with the other two (head-initial) outgroups, so the features are tracking typological similarity, not shared descent, at this depth. The study contributes 13 new treebanks for poorly documented languages, a reproducible framework for testing how much genealogical signal structural features carry, and direct evidence that, at Transeurasian time depths, this signal reflects typology rather than genealogy. Full article

Yellowfin tuna heads (YFTH), a major by-product of tuna processing, are rich in lipids and long-chain n-3 polyunsaturated fatty acids. This study developed and optimized an ultrasound-assisted extraction process without exogenous protease addition, termed non-enzymatic ultrasound-assisted extraction (NE-UAE), for recovering yellowfin tuna head oil (YFTO). The term NE-UAE distinguishes this process from the enzyme-assisted comparators and does not denote a different ultrasonic mechanism. A four-factor, three-level Box–Behnken design was used to optimize nominal ultrasonic power, heat-treatment temperature, liquid-to-solid ratio, and ultrasonic treatment time. The predicted local-optimum conditions were 58.3% of the 1000 W rated power (nominal setting, 583 W), 64.0 °C, 3.0 mL/g, and 31 min, respectively, yielding a validated oil recovery of 63.92 ± 0.24%. The NE-UAE oil (NETO) was compared at the process-chain level with oils obtained by papain-, trypsin-, and Alcalase-assisted aqueous extraction. Under the tested conditions, NETO had lower acid value, initial peroxide value, and p-anisidine value, a lighter color, and lower relative abundances of putative oxidation-related electronic-nose marker signals. Polyunsaturated fatty acids comprised 38.39% of total fatty acids in NETO, including 7.43% EPA and 27.13% DHA. During accelerated storage at 60 °C for 14 d, NETO showed lower absolute peroxide values and smaller baseline-corrected increases than the comparator oils. Overall, NE-UAE provided a feasible process option for recovering YFTO with acceptable oil recovery and favorable quality-related characteristics under the specific process conditions and comparator process chains examined in this study. Full article

Background: Asymmetries in head movement are common in postural disturbances. The correction of such changes using prismatic lenses has been reported but not quantified. Methods: A controlled randomized double-blind prospective study was performed on a population with Postural Deficiency Syndrome, comparing active head extension and rotation using prismatic and sham lenses. The evaluation was carried out using clinical photography and goniometer before intervention and 15 min after lenses’ application. Results: One hundred patients were included. The initial extension from a horizontal gaze was $53.2\pm 12.6$°, and the rotation amplitude was $113\pm 15.2$°. The overall head movement was below the standard reference values of 75° for head extension and 160° for rotation. ($p<0.01$ for both variables). The intervention group exhibited both an extension increase and rotation increase in amplitude: rotation improved by $8.7\pm 9.9$° ($p<0.01$), and extension increased on average $6.5\pm 8.0$° ($p<0.01$). The initial head rotation asymmetry was $10.1\pm 4.3$°, reducing to $1.70\pm 1.58$° after intervention ($p<0.01$). Conclusions: Patients with Postural Deficiency Syndrome have a limited and asymmetrical range of motion of head rotation and extension. These limitations are responsive to the use of low-powered prismatic lenses. Full article

Objectives: AC02, a novel synthetic peptide derived from ACTH, is being developed as a potential therapeutic alternative to porcine ACTH_1-39 for use in infantile spasms. Methods: The study comprised single-ascending dose cohorts (0.02, 0.04, 0.08, 0.16 mg/kg AC02) and multiple-ascending dose cohorts (0.04, 0.08 mg/kg AC02 daily for 5 days). A separate positive-control arm received porcine ACTH_1-39 (25 U/day for 5 days). PD effects (free and total cortisol) were compared head-to-head with the positive control. Plasma concentrations of AC02, porcine ACTH_1-39, free and total cortisol were quantified by validated LC-MS/MS methods, and anti-drug antibody responses were measured using a validated electrochemiluminescence bridging immunoassay. Population PK/PD modeling characterized the concentration–response relationships for free cortisol. Results: Across the 0.02–0.16 mg/kg dose range, AC02 exhibited approximately dose-proportional pharmacokinetics and no accumulation after multiple doses. At 0.04 and 0.08 mg/kg, the baseline-corrected effects on free cortisol were generally consistent with those of porcine ACTH_1-39. Dynamic monitoring of the free cortisol fraction revealed a biphasic pattern across all groups: an early peak followed by a later rise. Total cortisol, by contrast, showed only a monophasic decline, indicating the limitations of total cortisol as a sole PD biomarker. No unexpected safety signals were observed. At 0.04 mg/kg, AC02 demonstrated pharmacodynamic responses comparable to those of marketed ACTH product based on the biologically active component of free cortisol, with favorable safety and pharmacokinetic profiles. Conclusions: This first-in-human study demonstrates that AC02 has favorable pharmacokinetic properties, and an acceptable safety profile. Full article

Background: The video head impulse test (vHIT) is a standard tool for assessing semicircular canal function through vestibulo-ocular reflex (VOR) gain and corrective saccades (CS). Although reduced gain accompanied by CS typically indicates vestibular hypofunction, a paradoxical pattern—normal gain with CS—is occasionally observed, and its clinical significance remains unclear. Methods: We retrospectively analyzed 1573 patients who underwent both vHIT and bithermal caloric testing between September 2022 and September 2025. Among them, 1174 subjects with normal bilateral horizontal VOR gains (0.8–1.2) were evaluated for CS. CS were defined as at least three consistent refixation saccades on one side, verified visually. Subjects were categorized into Right-, Left-, and Bilateral-CS groups. Comparisons were made between CS(+) and CS(−) sides, as well as among groups, using paired and one-way ANOVA tests. Results: CS were identified in 46.3% (544 in 1174) of subjects with normal VOR gains, most frequently on the right side. The side showing CS exhibited significantly lower caloric responses than the contralateral side (p < 0.001). Bilateral CS were more common in older patients. Canal paresis values indicated that the CS side corresponded to the weaker caloric side, suggesting subtle unilateral hypofunction. Conclusions: Even with normal gain, the presence of CS may reflect mild vestibular asymmetry rather than a benign artifact. In addition, bilateral CS were associated with older age, suggesting a possible contribution of age-related vestibular or compensatory mechanisms. Careful control of methodological biases and integration with other vestibular tests are essential for interpretation. Full article

Accurate positioning remains a central challenge in head and neck radiotherapy because highly conformal dose distributions are delivered near multiple critical organs, and clinically relevant discrepancies may not be resolved by rigid image-guided correction alone. This narrative review synthesizes evidence on positioning management in head and neck radiotherapy, with particular emphasis on the transition from couch correction to action-oriented clinical decision-making. Peer-reviewed studies, guidelines, and key reviews published between January 2005 and April 2026 were identified using PubMed/MEDLINE and reference screening. Positioning uncertainty is often region-specific and dynamic, reflecting lower-neck and shoulder mismatch, mandibular variation, mask-fit deterioration, functional motion, and progressive anatomical changes during treatment. Thermoplastic immobilization, customized headrests, bite blocks, cone-beam computed tomography, six-degree-of-freedom correction, and adaptive workflows have improved reproducibility and verification; however, these approaches do not eliminate mismatches caused by non-rigid posture, immobilization failure, or evolving anatomy. In such situations, the key clinical task is to distinguish rigid setup deviations suitable for online correction from discrepancies requiring repositioning, re-immobilization, repeat simulation, or adaptive replanning. Emerging technologies, including artificial intelligence, dose-of-the-day estimation, corrected CBCT, synthetic CT, rapid replanning, and workflow-integrated decision support, may help close the gap between mismatch detection and intervention selection. Positioning management should therefore evolve from a correction-centered model to an action-oriented framework that links observed geometric or anatomical change to the most appropriate clinical response. Full article

In 3D integrated circuits, a placement decision that looks acceptable from a 2D wirelength view can still create a local thermal or stress problem after stacking. This issue becomes more visible as the number of tiers and the density of vertical interconnects increase. We propose ThermIC, a placement framework that brings thermal and mechanical risk estimates into the placement loop rather than treating them only as post-layout checks. The novelty of ThermIC does not lie in treating graph neural networks, reinforcement learning, uncertainty-aware learning, or physics-informed regularization as individually new techniques. Instead, ThermIC contributes a placement-time coupling mechanism in which physically typed graph propagation, dense multi-constraint risk prediction, and action-level reinforcement learning feedback are jointly organized for stacked 3D-IC placement. ThermIC uses a heterogeneous graph encoder to carry thermal, stress, timing, and congestion information through the netlist; a constraint head to estimate local hotspot, stress-risk, timing-violation, and congestion probabilities; and a sequential placement policy trained with physics-informed penalties. We evaluate the method on ThermIC-Bench, a simulated corpus with more than 30,000 finite-element samples from 18 heterogeneous 3D-IC designs with 4–8 tiers. Because the present study does not include proprietary industrial circuits, silicon measurements, or a tape-out case, the experimental results are interpreted as simulation-based benchmark evidence rather than final industrial qualification. ThermIC connects the heat-kernel branch to the discretized heat-conduction equation and the stress-filter branch to linear thermo-elastic equilibrium, providing a mechanism-level basis for physical interpretability. The analysis distinguishes offline simulation/training cost from online deployment cost and reports complexity, runtime, and memory scaling for practical large-scale use. Under joint DRC, thermo-mechanical stress, and thermally coupled timing checks, ThermIC obtains an $82.1\%$ physical verification pass rate. The peak-temperature error is $3.1$ °C, the hotspot localization IoU is 0.89, and the number of placement-closure iterations is reduced by $3.7\times$ relative to the heuristic baseline. Together, these benchmark results indicate that early, differentiable multi-physics feedback can make 3D placement less dependent on late correction cycles. Full article

Background/Objectives: Colorectal cancer (CRC) is the second leading cause of cancer-related mortality worldwide, with polyp miss rates of up to 26% reported during colonoscopy and classification accuracy remaining highly operator-dependent. Accurate multi-class polyp subtype classification is clinically critical, as it directly determines treatment decisions: adenomatous polyps require resection, whereas hyperplastic lesions may warrant only surveillance. This study aims to systematically compare five deep learning architectures for five-class gastrointestinal polyp classification and to provide clinically interpretable diagnostic insights through Grad-CAM visualisation. Methods: ResNet50, VGG16, EfficientNet-B3, DenseNet121, and Vision Transformer (ViT-B/16) were evaluated on the Kvasir Dataset V2 (5000 images, five classes) under a unified training and evaluation protocol on common GPU hardware. All models employed ImageNet transfer learning with a redesigned multi-layer classification head. Two optimisation strategies were applied: SGD with cosine annealing for CNN architectures, and AdamW with linear warmup for ViT-B/16. Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to generate spatial attention heatmaps for qualitative clinical interpretation. Results: Under a single 80/10/10 split, ViT-B/16 attained the highest accuracy (97.2%); however, because a single split is sensitive to sampling, the evaluation was strengthened with stratified five-fold cross-validation (mean ± SD). Under cross-validation, EfficientNet-B3 achieved the highest accuracy at 95.90 ± 0.35%, followed closely by ViT-B/16 (95.12 ± 0.72%), then ResNet50 (91.74 ± 0.74%), DenseNet121 (90.32 ± 0.70%), and VGG16 (88.50 ± 1.72%); the small standard deviations indicate that all models, including ViT-B/16, were stable across folds. Pairwise McNemar tests with Holm correction found that every difference was statistically significant (p < 0.05), including the EfficientNet-B3 advantage over ViT-B/16 (p = 0.010). ViT-B/16 thus remained a strong, stable performer that significantly outperformed the three remaining CNNs, while the cross-validated ranking placed the most compact model, EfficientNet-B3, first: a Vision Transformer was highly competitive with, but not superior to, the strongest CNN. A consistent, architecture-agnostic misclassification pattern was identified between dyed-lifted polyps and dyed-resection margins across all five models, consistent with a task-level visual ambiguity that may also reflect overlapping class definitions and annotation factors, with direct clinical implications. Grad-CAM analysis, quantified by attention-entropy and concentration metrics, showed that model attention remained focused on relevant stained tissue regardless of whether predictions were correct, indicating that the dyed-class confusions reflect genuine visual ambiguity rather than a localisation failure. Conclusions: Under cross-validation, EfficientNet-B3 achieved the highest accuracy on the Kvasir V2 five-class task, significantly outperforming all other architectures, with ViT-B/16 being a close and competitive second. The identified confusion between post-procedural chromoendoscopic classes is unlikely to be fully resolved by architectural changes alone and may require higher-resolution imaging or domain expert re-annotation. These findings contribute to the evidence base for explainable deep learning in gastrointestinal endoscopy; external, multi-centre validation remains necessary before clinical adoption. Full article

Staged upward seepage can trigger particle migration and pore-structure adjustment in graded sandy soils, but the resulting transition behavior remains difficult to identify quantitatively. In this study, three representative sandy soils from a deep overburden deposit in southeastern Tibet were tested using a laboratory vertical upward seepage apparatus. Eight specimens with different nominal preparation states were subjected to stepwise increases in hydraulic head difference. Local hydraulic gradient, seepage velocity, hydraulic conductivity, and macroscopic outflow phenomena were monitored. Apparent seepage-transition hydraulic gradients were identified by combining abrupt changes in k–i curves, conductivity ratios between eligible staged records, and observed seepage responses. The clearest transition occurred in the nominal loose specimen of Soil 2, where the temperature-corrected hydraulic conductivity k₂₀ increased from 1.76 × 10⁻³ to 2.25 × 10⁻² cm s⁻¹ as i increased from 0.20 to 0.25, giving k_20,2/k_20,1 = 12.80 and i_c = 0.225. A clear transition was also identified for the nominal dense specimen of Soil 3, with k_20,2/k_20,1 = 6.61 and i_c = 0.583. Clear transitions were identified in the tested specimens only for Groups D and H, whereas the remaining specimens showed weak or phenomenon-assisted responses, local high-gradient fluctuations, or anomalous loading-path records rather than uniformly identifiable transition points. These results show that apparent transition gradients are path-dependent and should be evaluated together with loading history, seepage-velocity evolution, conductivity ratios, and macroscopic observations. Full article

Background: HER2 status guides targeted therapy in breast cancer but is currently determined by invasive biopsy. Imaging-based HER2 prediction from dynamic contrast-enhanced MRI (DCE-MRI) could provide a non-invasive adjunct decision-support signal, but published models are typically single-center with heterogeneous preprocessing that limits reproducibility. Methods: We trained a Triple-Head Dual-Attention ResNet (THDA-ResNet) that processes three DCE phases (pre-contrast, early post-contrast, and late post-contrast) on the multicenter BreastDCEDL dataset (n = 1149, I-SPY trials), and we compared it with Vision Transformer (ViT) and Convolutional Vision Transformer (CvT) baselines, all ImageNet-pretrained. We benchmarked 14 preprocessing strategies, with and without N4 bias-field correction. External validation used the independent BreastDCEDL_AMBL cohort (43 lesions). AUC confidence intervals used stratified bootstrap; model comparisons used DeLong’s test. Results: THDA-ResNet achieved the highest AUC, 0.74 (95% CI 0.65–0.83), versus 0.66 for ViT and 0.63 for CvT, with the advantage reaching borderline significance over CvT ($p=0.054$) and not significant over ViT ($p=0.14$). At a threshold of 0.7, it retained discrimination (sensitivity 0.41, specificity 0.86), while transformers collapsed to near-trivial classifiers. External AUC was 0.66 (0.49–0.81). N4 correction did not improve performance. Conclusions: Attention-driven CNNs are a strong default for HER2 prediction from DCE-MRI on medium-sized cohorts, and N4 correction can be omitted, simplifying the pipeline. Full article