IS&T | Library

Abstract

The categorization of medical images is essential for disease screening, diagnostic processes, and the design of therapeutic interventions. Existing methods mainly extract image features through a single spatial domain approach. However, medical images are characterized by significant noise interference, small interclass differences, and large intraclass differences, which result in relatively low classification accuracy. Therefore, this work designs a dual-path feature extraction mechanism that synergistically complements frequency and spatial domains as well as local and global features. It combines the strength of convolutional neural networks in extracting local features, the capacity of Transformers for global modeling, and the multi-scale analysis of discrete wavelet transform (DWT) to extract rich medical image features and improve classification accuracy. The model is mainly implemented through a dual-branch architecture. One branch uses a lightweight VGGNet to efficiently extract hierarchical local features and capture subtle pathological manifestations. The other branch combines DWT with the vision Transformer (ViT). First, DWT is used to perform multi-scale frequency domain decomposition on the image, where low frequencies retain organ morphology and lesion contours and high frequencies capture tissue microstructure and lesion boundaries, enhancing feature expression while suppressing noise. Then the ViT is used to establish cross-regional global semantic associations and model long-term dependencies. The features from the two branches are integrated through an adaptive fusion mechanism to form a comprehensive and robust representation. This method achieves an accuracy of 0.93 and an AUC of 0.94 on the public dataset Kaggle-EDC, with both sensitivity and specificity outperforming current mainstream models. Experimental results demonstrate that this framework markedly enhances both the classification performance and robustness on medical images, providing an effective solution for precise medical image analysis.

Digital Library: JIST

FastTrack Article

A Cascaded-XCViT based Few-shot Classification Algorithm for Breast Ultrasound Images

19 0

breast ultrasound
few-shot learning
vision transformer
self-supervised learning
cascaded networks

Tong Ying, Zhao Xiaoyan, Jiang Zhuoyu, Chen Guofeng

Pages 1 - 10, © Society for Imaging Science and Technology 2027

DOI

10.2352/J.ImagingSci.Technol.2027.71.2.020501

Abstract

Breast ultrasound serves as a cornerstone non-invasive technique for the screening and diagnosis of breast diseases. To address the challenge of limited feature generalization in few-shot learning for breast ultrasound image classification, this study proposes a novel cascaded Cross-Covariance Vision Transformer (XCViT) framework. A dual-stage architecture is introduced. Stage A employs hierarchical feature distillation via a three-level cascaded-XCViT module with Teacher–Student frameworks, generating dual-path tokens and utilizing token spectral clustering for semantic downsampling. Stage B uses frozen encoder outputs to derive prototype vectors from support-set features, enabling training-free inference via distance metric learning. Evaluation showed accuracies of 79.8% (1-shot) and 89.7% (5-shot) on CIFAR-FS, 47.2% (1-shot) and 67.3% (5-shot) on FC100, and 61.5% (5-shot) on a clinical breast ultrasound dataset, outperforming existing methods by 2.0–5.6%. The method demonstrates strong performance in low-data regimes and offers a promising solution for few-shot clinical auxiliary diagnosis.

Digital Library: JIST

FastTrack Article

Trimming Path Detection of Invisible Orthodontic Aligners based on an Iteratively Approaching Gingival Margin

14 0

invisible orthodontic aligners
trimming path
gingival margin
digital dentistry
curvature
3D models

Rui-Bin Chern, Tzung-Han Lin, Shyh-Yuan Lee

Pages 1 - 9, © Society for Imaging Science and Technology 2027

DOI

10.2352/J.ImagingSci.Technol.2027.71.2.020502

Design of an Automated Pallet Docking Robot System with Depth Image Detection and Pose Recognition

Abstract

The growing demand for aesthetic orthodontic solutions and the increasing clinical use of clear aligners have created a need for robust digital methods that improve production efficiency and assist clinicians in defining accurate aligner boundaries. This study presents a practical method for determining the trimming paths of clear digital orthodontic aligners by iteratively detecting gingival margin lines on three-dimensional dental models. The proposed approach integrates pose adjustment, curvature estimation, and clustering to delineate gingival margins, which are critical for precise aligner manufacturing and patient comfort. Principal-curvature-based filtering is first employed to identify candidate gingival margin vertices, followed by clustering to remove spurious regions with similar curvature patterns on the crown and gingival surfaces. An iterative contour approach scheme then refines the candidate trimming paths by projecting an initial enclosing contour, derived from a statistical analysis of multiple tooth models, onto each 3D dental model under different approach rates. Experiments on multiple tooth models demonstrate that the proposed method consistently generates clinically acceptable trimming paths and exhibits adaptability across various dental morphologies. Parametric analysis shows that an acceleration term between 0.4 and 0.6 yields stable contour shrinkage, minimizes path misalignment, and enhances smoothness, thereby improving aligner retention and patient comfort. The proposed framework provides a practical, reproducible solution for the digital generation of trimming paths for invisible orthodontic aligners, supporting high-quality aligner production and enabling quantitative assessment of gingival margins in digital dentistry workflows.

Digital Library: JIST

FastTrack Article

IMETI 2025 Special Issue

18 0

forklift mobile robots
YOLOv8
pallet docking system
pallet detection
pallet pose estimation

Ching-Chang Wong, Zi-You Huang, Hsuan-Ming Feng, Chi-Yi Tsai

Pages 1 - 17, © Society for Imaging Science and Technology 2026

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050404

Abstract

This paper integrates deep learning technology, object detection, pose estimation, and map-based path navigation to design and implement an efficient pallet automatic docking robotic system. To fulfill the stringent requirements of high-precision pallet docking in indoor environments, this study presents a novel perception-to-action navigation framework that tightly integrates perception, localization, planning, and control. In the perception stage, a real-time pallet detection and instance segmentation method based on YOLOv8-Seg is proposed. The segmentation results are further combined with the Random Sample Consensus algorithm and Principal Component Analysis to improve the robustness and accuracy of pallet pose estimation. In the navigation stage, the proposed framework seamlessly switches between a macro-navigation phase and a micro-alignment phase. During macro-navigation, a SLAM-generated occupancy map is employed for environmental modeling, Dijkstra’s algorithm is utilized for global path planning, and a localization framework combining Adaptive Monte Carlo Localization and the Extended Kalman Filter is adopted to mitigate cumulative localization drift. Moreover, the Dynamic Window Approach provides real-time local obstacle avoidance and motion control. During micro-alignment, a vision-based Dynamic Visual Servoing scheme continuously incorporates pallet pose priors estimated by the perception module into the local control loop. This enables the mobile robot to actively compensate for lateral and angular errors during the final docking maneuver, thereby ensuring centimeter-level localization accuracy and robust pallet engagement performance. The effectiveness of the proposed framework is validated through rigorous physical experiments, demonstrating its ability to achieve superior docking accuracy and smooth material transportation.

Digital Library: JIST

FastTrack Article

Quantitative Measurement of Binocular Just-Noticeable Color Difference in the CIE 1976 u′v′ Chromaticity Diagram for Near-Eye Display Applications

30 1

just-noticeable color difference
binocular color perception
near-eye display
CIE 1976 u′v′ chromaticity diagram
luminance contrast

Zheng Huang, Sheng Lin, Zhiyu Chen, Hanwen Gong, Chajuan Liu, Qiang Liu

Pages 1 - 15, © Society for Imaging Science and Technology 2026

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050509

Abstract

The color consistency between left and right eye images impacts users’ overall visual experience when viewing near-eye displays (NEDs). This has received insufficient attention in the literature. In this study, experiments were conducted to measure the binocular just-noticeable color difference (JNCD) from the CIE 1976 u′v′ chromaticity diagram. A custom binocular viewing apparatus, which contains no lenses, prisms, or mirrors, was developed to allow for accurate and reliable measurement of the binocular JNCD. Five color stimuli (neutral, magenta, green, blue, and yellow) at three luminance levels (100, 150, and 200 cd/m²) were adopted as the reference stimuli in a series of adjustment experiments. During these experiments, observers were required to determine the binocular JNCD across eight color-change directions under three background luminance conditions (0, 40, and 80 cd/m²). It was found that the binocular JNCD, as measured on the u′v′ diagram, was significantly influenced by the direction of color change and the luminance contrast between the stimulus and the background. Specified by the one-standard-deviation ellipse, the experimental data indicated that the binocular JNCD values correspond to 0.0024, 0.0044, 0.0043, 0.0040, and 0.0038 u′v′ units for neutral, magenta, green, blue, and yellow colors, respectively. These findings should contribute to a deeper understanding of binocular color perception and provide a useful reference for NED color calibration and industrial optical design.

Digital Library: JIST

FastTrack Article

Fair and Robust Federated Learning for Cardiovascular Disease Prediction via Client Heterogeneity Quantification

22 0

federated learning
non-IID data
fairness
cardiovascular disease

Xiaolu Xu, Yulong Li, Shuai Zheng, Jia Shang, Chengjie Lu, Hongbin Lu, Yonggong Ren, Qiannan Zhang

Pages 1 - 13, © Society for Imaging Science and Technology 2026

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050507

Abstract

Cardiovascular disease (CVD) remains a major global health burden and the leading cause of mortality worldwide. Although machine learning has shown promise for CVD risk prediction, centralized training on sensitive medical data from multiple hospitals raises serious privacy and security concerns. To address this issue, we propose FedCH, a novel federated learning method for collaborative CVD prediction. Specifically, FedCH quantifies client heterogeneity by measuring the similarity between local and global models together with client-specific prediction accuracy. These metrics are used to dynamically weight the global aggregation process, thereby improving the overall model performance while promoting fairness across clients. In addition, by incorporating historical global model parameters into the update process, FedCH mitigates local model fluctuations and enhances the generalization and robustness of the global model. Comprehensive evaluations on four public datasets and a real-world cardiovascular dataset from a tertiary hospital in China show that FedCH consistently outperforms five state-of-the-art federated learning baselines under heterogeneous client settings.

Digital Library: JIST

FastTrack Article

Color Characterization of a Polarization Camera in Comparison with a Conventional Color Camera

31 5

colorimetric characterization
polynomial regression
polarization camera
degree of linear polarization (DoLP)
chromatic shift

Yu-Chun Hung, Tzung-Han Lin

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050508

Deep Neural Networks with Local Texture and Geometric Augmentation for Facial Expression Recognition

Abstract

In this study, the color performance of a polarization camera was investigated and compared with that of a color camera. Three experiments were conducted using a well-calibrated imaging system with a 45∘∕0∘ CIE configuration. First, among the seven preset equations, the 3 × 11 characterization matrix exhibited a small regression error. Second, the use of unpolarized and polarized light led to different chromatic shifts, particularly for sections with 90∘ and 0∘ polarization. Lightness (L^∗) was found to be more closely related to the degree of linear polarization (DoLP), whereas chromaticity showed less correlation. Finally, the impact of the two levels of glossy samples was examined using both the cameras. The surface characteristics of the objects were observed and correlated with variations in the DoLP and chromatic shifts. This work analyzes the color characteristics of a polarization camera from FLIR. These findings contribute to understanding the influence of polarized light sources, color characterization conversions, and sample variations on camera color accuracy.

Digital Library: JIST

FastTrack Article

IMETI 2025 Special Issue

23 2

facial expression recognition
geometric augmentation
local binary patterns
channel attention mechanism

Zi-Jie Lin, Ching-Yi Chen

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050403

Abstract

Facial expression recognition (FER) under unconstrained environments faces two severe challenges: class imbalance and the loss of high-frequency texture information. To address these challenges, this study proposes TexGeo-Net, a dual-stream deep architecture that effectively rebalances the training data distribution to mitigate the long-tail effect by employing a landmark-guided geometric augmentation strategy while simultaneously utilizing local binary patterns to precisely preserve micro-texture details. In terms of model design, the system adopts the Densely Connected Convolutional Network as the backbone and deeply integrates the Squeeze-and-Excitation channel attention mechanism. Through adaptive feature fusion, this design compensates for the spatial information loss caused by downsampling in deep networks. Experimental results on three benchmark datasets—CK+, RAF-DB, and FER2013—demonstrate that TexGeo-Net performs competitively with existing methods in terms of recognition accuracy and robustness across these heterogeneous datasets. This validates the complementary advantages of geometric structures and texture features, demonstrating the feasibility of TexGeo-Net in real-world applications.

Digital Library: JIST

FastTrack Article

General Domain Event Detection Method based on TAP-BERT and Event Fusion

22 1

event detection
TAP-BERT
attention mechanism
CRF

Xinyu He, Manfei Kan, Yuning Zhang, Yonggong Ren

DOI

10.2352/J.ImagingSci.Technol.2026.70.5.050506

Road-Distress Detection Algorithm for Low-Illumination and Blurred Images based on Image Enhancement

Abstract

Event detection is a key task in information extraction that aims to identify event trigger words from text and classify them into event types. Existing methods focus more on the trigger word itself, overlooking the dependencies between contexts and lacking comprehensive feature extraction. Therefore, this paper proposes the model TBEF (TAP-BERT + Event Fusion Layer + CRF). First, the TAP-BERT pretrained model for word vector embedding—which incorporates word embedding matrix decomposition, feed-forward neural network pruning, and self-attention mechanism pooling in BERT—is utilized to obtain richer contextual semantic information. Then an event fusion layer consisting of attention and gate mechanisms is employed to better compute the correlation between events and trigger words. Finally, the sequences are labeled by Conditional Random Field (CRF) integrating BIO labeling. Experimental results show that the proposed method yields an F₁ value of 68.55% on the MAVEN dataset, achieving advanced performance for event detection.

Digital Library: JIST

FastTrack Article

IPEC Special Issue

23 0

low-illumination blur
Retinex
road defects
YOLOv8
lightweight

Huiying Yin, Yu Zhang, Liu Yang, Heng Liu, Lijun Ma, Xinhao Zhao

DOI

10.2352/J.ImagingSci.Technol.2026.70.6.060401