Background: Cranial nerve enhancement is traditionally assessed with T1-weighted imaging, but heavily T2-weighted 3D FLAIR (HT2-FLAIR) imaging offers greater sensitivity to subtle T1 shortening. This study aimed to assess contrast enhancement of the oculomotor, trigeminal, and facial nerves using HT2-FLAIR imaging. Methods: This retrospective repeated-measures cohort study included 33 patients. We subjectively assessed contrast enhancement of the cisternal segments of these three cranial nerves across precontrast, postcontrast, and 4 h delayed postcontrast HT2-FLAIR images and objectively evaluated serial changes in the signal intensity ratio (SIR)—the ratio of each structure’s signal intensity to that of the pons—for the oculomotor and trigeminal nerves, cerebrospinal fluid (CSF) spaces, and choroid plexus. Results: Subjectively, a high rate of enhancement (>90%) was observed bilaterally for all three nerves on 4 h delayed postcontrast images. Objectively, SIR of both the oculomotor and trigeminal nerves peaked on 4 h delayed postcontrast images, followed by postcontrast images, paralleling the pattern of CSF spaces and differing from that of choroid plexus. Conclusions: This study is the first to show that HT2-FLAIR imaging can visualize temporal enhancement of the cisternal segments of cranial nerves. Extravascular gadolinium may represent an alternative mechanism for normal cranial nerve enhancement.
To address the electrical safety hazards caused by missing zero-value insulators in high-voltage transmission lines, this study proposes a complementary detection method based on the fusion of infrared images and electric field intensity features. Infrared images of insulator strings are acquired using a UAV-mounted infrared thermal camera, and the surface temperature of each insulator is extracted via YOLOv8 to construct a temperature anomaly factor. Meanwhile, a three-dimensional electric field model of the insulator string is established using COMSOL Multiphysics 6.2 to extract the axial electric field intensity and calculate the electric field deviation. These two types of features are combined into a two-dimensional vector, and the Minimum Covariance Determinant (MCD) is used to estimate robust mean and covariance, from which a robust Mahalanobis distance is calculated to generate a fusion score for zero-value insulator identification. A three-level classification criterion is established, categorizing insulators as normal, suspicious, or zero-value. Experimental results demonstrate that this method effectively exploits the complementarity of infrared and electric field features, achieving the accurate identification of zero-value insulators, with a precision of 93.8%, zero false-negative and false-positive rates, and a suspicious rate of 6.2%. The method is validated to be highly accurate and robust.
Standard convolution samples features on a fixed grid whose locations do not adapt to the input content. We develop the Geometry-Adaptive Sampling and Hierarchical Feature-Sifting Network (GeoSieveNet) based on a 34-layer residual network (ResNet-34). The rotational deformable convolution module (RDCM) assigns predefined offsets with different orientations to individual channel groups and combines them with input-predicted dynamic offsets to adjust sampling locations. Its feature preservation module (FPM) fuses responses from the fixed-grid and rotational deformable branches. Hierarchical gated attention (HGA) combines directional statistics with local convolutional responses to generate spatial gating weights within each channel group. GeoSieveNet achieved classification accuracies of 96.95%, 81.86%, 94.71%, and 89.79% on CIFAR-10, CIFAR-100, Imagenette, and Imagewoof, respectively. Across three independent runs, these results exceeded ResNet-34 under the same training settings by 0.66, 3.06, 4.73, and 5.19 percentage points, respectively. All paired tests were significant at the 0.001 level. Ablation experiments attributed gains of 0.36–1.89 percentage points to FPM alone, with larger gains on datasets exhibiting greater geometric variation. At the same floating-point operation budget as ResNet-34, GeoSieveNet retained an accuracy advantage of 0.54–3.96 percentage points.
This study analyzed how the performance of various neural network architectures responds to offline image augmentation using standard geometric and photometric transformations. We evaluated a composite training-data enrichment protocol using two multiclass color image classification tasks: a dataset of leg photographs classified according to the CEAP system (sourced from open-source repositories; 20,971 images, seven classes) and a dermatological image dataset (6786 images, eight classes). To enrich the initially balanced subsets, transformed images were added to classes underrepresented in the original datasets, while additional real images were added to the remaining classes. Twelve architectures were investigated, including classical CNNs, vision transformers, and a modernized CNN. Across the 12 architectures, 11 augmentation levels, and 20 runs per configuration, mean accuracy ranged from 0.55 to 0.68 (leg photographs) and from 0.59 to 0.79 (dermatological images). The expansion levels yielding the highest observed mean accuracy varied across configurations, with descriptive maxima occurring at levels from 0% to 100%. For each model–dataset configuration, the highest mean accuracy across the eleven levels was treated as a descriptive maximum. Standard deviations across runs characterized the variability of the training process, whereas standard errors reflected the uncertainty of mean estimates given a fixed test set. Transformer-based configurations demonstrated lower average run-to-run standard deviations; however, this tendency was configuration- and dataset-dependent, warranting further research. An exploratory augmentation-as-biased-variance-reduction (ABVR) model is proposed to describe the observed accuracy curves. Its fitted coefficients and maximizers summarize the available data and have not been independently validated.
Due to the limited availability of annotated medical datasets, medical image classification typically relies on transfer learning from large-scale natural image datasets such as ImageNet. However, it remains unclear when domain-specific pretraining offers greater benefits. This study evaluates the impact of different pretrained weights on neural network performance in radiological image classification tasks. Four strategies were compared: ImageNet, RadiologyNET, and RadImageNet model fine-tuning against learning from randomly initialised weights (Baseline). Experiments were conducted over ten seed-paired runs using ResNet50, InceptionV3, DenseNet121, EfficientNet-B3, and MobileNetV3-Large on three classification tasks: multiclass Chest CT, multiclass BRISC brain MRI, and binary FracAtlas X-ray fracture classification. Pretraining generally improved performance over the Baseline, with the largest improvements observed on the Chest CT task. The Friedman tests identified significant differences between initialisation strategies for every architecture on all three tasks (all p≤0.0002). ImageNet achieved the best results in most architecture–task combinations, while RadiologyNET was comparable or superior in selected cases, including three of five architectures on Chest CT. RadImageNet showed less consistent transfer, particularly with DenseNet121. Pretrained models also generally reached high validation performance more rapidly than randomly initialised models. These results further support previous findings that ImageNet-pretrained models can perform comparably to or better than domain-specific radiological models when fine-tuned on sufficient downstream data, while the observed differences are consistent with possible effects of source–target alignment, downstream data availability, and network architecture.
