International Journal of Imaging Systems and Technology最新文献

Cervical cancer is one of the most common malignant tumors among women worldwide, and accurate early diagnosis is critical for improving patient survival rates. Traditional cytological screening methods rely on manual microscopic examination, which suffers from low efficiency and high subjectivity. In recent years, deep learning has facilitated the automation of cervical cell image analysis, yet challenges such as insufficient modeling of pathological features and high computational cost remain. To address these issues, this study proposes a novel dual-branch multi-scale model, MorphoFormer. The model employs a multi-scale dilated Transformer (DilateFormer) as its backbone and innovatively incorporates specialized modules for each branch: A Local Context Aggregation (LCA) module in the local branch and a Global Focus Attention (GFA) module in the global branch. These modules respectively enhance the representation of local details and global semantics, and their features are fused to enable collaborative multi-scale information modeling. Experimental results on the publicly available SIPaKMeD dataset demonstrate that MorphoFormer achieves classification accuracies of 99.58%, 98.51%, and 98.14% for binary, three-class, and five-class tasks, respectively. Further validation on the Blood Cell Count and Detection (BCCD) dataset indicates strong cross-task robustness. Moreover, MorphoFormer requires only 8.22 GFLOPs for inference, highlighting its practical potential by achieving high performance with low computational overhead. Related codes: https://github.com/sijhb/MorphoFormer.

Traditional unsupervised domain adaptation methods usually depend on source domain data distribution for cross-domain alignment. However, direct access to source data is often restricted due to privacy concerns and intellectual property rights. Without using source data, Source-free unsupervised domain adaptation methods can align the pre-trained model with the target domain by generating pseudo-labels for target domain data, which are then used as labeled samples to guide transfer learning. However, methods that generate pseudo-labels solely through iterative averaging often neglect spatial correlations among pixels and are susceptible to noise, resulting in blurred label boundaries. To this end, we propose a source-free domain adaptation framework for fundus image segmentation, which consists of a Multiscale Feature Fusion module for generating high-quality pseudo-labels and a Stepwise Attention Integration module for enhancing model training. The Multiscale Feature Fusion module refines the initial pseudo-labels from the pre-trained model through neighborhood value filling, effectively reducing noise and sharpening label boundaries. The Stepwise Attention Integration module progressively integrates high-level and low-level feature information into the low-level representation. The fused features preserve high-resolution details and enrich semantic content, thereby substantially enhancing the model's recognition capability. Experimental results demonstrate that, without using any source domain images or modifying the pre-trained model, our method achieves performance comparable to or even surpassing state-of-the-art approaches.

Retinopathy, a prevalent retinal disorder, poses a major risk of vision loss if not detected at an early stage. Automatic lesion segmentation plays a key role in effective diagnosis and disease monitoring. In this work, we present a complete and interpretable pipeline that combines YOLOv12 for lesion segmentation, SVD-CAM for visual explanation, and transformer-based Gradient Boosted Neural Networks (GBNN) and Random Forest classifiers for diabetic retinopathy (DR) severity grading. YOLOv12 (You Only Look Once, version 12), known for its real-time object detection capability, is adapted to the complex task of retinal lesion segmentation, delivering both high accuracy and speed. To enhance lesion localization, SVD-CAM generates precise heatmaps that highlight critical pathological regions influencing the grading decision. The segmented lesions are then quantified and used as input features for the grading stage, enabling clinically aligned DR classification. Our approach not only achieves state-of-the-art performance across three public datasets (IDRiD, DDR, and FGADR) but also provides lesion-level interpretability that improves clinical trust and adoption. Extensive experiments demonstrate that the proposed framework delivers accurate segmentation, reliable grading, and meaningful visual explanations, establishing a robust solution for automated DR analysis.

Accurate identification and segmentation of teeth in cone-beam computed tomography (CBCT) images are essential for dental diagnosis and treatment in digital dentistry. However, extracting regions of interest (ROI) from maxillofacial CBCT images remains difficult due to the low pixel ratio of tooth structures, especially in the apical area. Traditional tooth segmentation methods such as threshold-based, region-based, and edge-based methods address limited accuracy under challenging imaging conditions. In this paper, we propose TTNNet, a three-stage tooth instance segmentation network designed to improve tooth segmentation from CBCT images. The proposed TTNet employs an intersection-based refinement of the tooth centroid heatmap, retaining only pixels that simultaneously lie within the predicted tooth mask and high-probability heatmap regions. This intersection operation eliminates noise in the tooth centroid heatmap, such as regions that may have been labeled as teeth incorrectly. Extensive experiments on publicly available CBCT tooth dataset demonstrate that TTNet achieves superior performance compared to recent state-of-the-art methods.

The rising prevalence of lumbar spine disorders demands scalable solutions for mass screening and automated diagnosis. Accurate analysis of specific MRI slices, such as mid-sagittal or transverse mid-height intervertebral disc (IVD) slices, is essential but currently relies on time-consuming, error-prone manual selection. Automating this process is crucial to enhance the efficiency and accuracy of computer-aided diagnostic systems. To address this need, this study introduces a novel deep learning-based framework—SpineDeep-Net that integrates self-attention mechanisms within a multi-layer convolutional neural network for automatic selection of optimal transverse planes of lumbar spine MRI disc slices. By focusing on mid-height slices of L3/L4, L4/L5, and L5/S1 IVDs—the most diagnostically relevant slices, SpineDeep-Net eliminates the reliance on manual selection processes, thereby accelerating and improving the diagnostic pipeline. Unlike standard attention, the proposed dual-self-attention employs two sequential attention stages that jointly enhance long-range spatial cue extraction and emphasize subtle disc-level differences. This mechanism enables the model to focus more effectively on diagnostically relevant regions within lumbar MRI slices by dynamically recalibrating feature maps and strengthening feature dependencies. Experimental evaluations demonstrate the superior performance of SpineDeep-Net, achieving 96.83% accuracy and 98.41% specificity, outperforming state-of-the-art methods. By automating the selection and classification of clinically critical disc slices, SpineDeep-Net addresses a key challenge in lumbar spine diagnostics, providing a reliable, scalable, and efficient tool that aids radiologists in making informed clinical decisions. The proposed framework highlights the transformative potential of self-attention-guided deep learning in advancing healthcare diagnostics. The source code is publicly available at https://github.com/rakeshchandrajoshi/spinedeepnet.

Whole-slide imaging assisted by deep learning has been employed to help the digital pathology, while limited by the scarcity of paired label data. To address this issue, a novel self-supervised image modeling framework, PathMAE, is proposed to effectively enlarge the labeled dataset in a cross-domain way, where cross-dataset and even cross-disease histopathological images can be used for model training. PathMAE integrates masked image modeling and contrastive learning to effectively learn transferable visual representations from unlabeled WSIs. The framework comprises two key components: a Swin-Transformer-based encoder-decoder (SMED) with a window-masking strategy for local feature reconstruction, and a Dynamic Memory Contrastive Learning (DMCL) module for enhancing global semantic alignment via memory-guided feature comparison. Experimental results on three public histopathology datasets demonstrate the robustness and generalizability of the proposed method. In cross-disease transfer (BreakHis → Osteosarcoma), PathMAE achieved 97.15% accuracy and 99.03% AUC; in cross-dataset transfer (BreakHis → Camelyon16), it obtained 84.67% accuracy and 88.04% AUC. These findings validate the capability of PathMAE as a scalable and domain-adaptive image analysis framework, offering new potential for building reliable computational pathology systems under limited supervision.

This study focuses on iris recognition using deep learning techniques. The EfficientDet network model was employed for both iris detection and recognition tasks. Four datasets were utilized to train and evaluate the deep learning network. The model was trained to extract iris features and classify individuals based on their unique iris patterns. The proposed method achieved a high recognition rate of over 98% across multiple dataset evaluations. For real-time implementation on an embedded system, the trained model was quantized to an 8-bit integer format to accommodate resource-constrained devices. Despite this quantization, the recognition accuracy remained high, reaching 97%. By incorporating an Edge TPU accelerator alongside a Raspberry Pi system, the processing speed reached up to 10 frames per second during real-time iris camera testing, demonstrating the feasibility of real-time iris recognition. An intruder test was conducted to assess the system's robustness in preventing unauthorized access. The False Acceptance Rate (FAR) was measured to assess the likelihood of incorrectly accepting an unauthorized individual. Experimental results show that the FAR can be reduced to zero by applying additional temporal constraints, effectively preventing unauthorized individuals from passing the iris recognition-based access control system.

An ultrafast two-dimensional (2D) imaging system with large field of view is proposed by leveraging virtually imaged phased array through Galilean beam expanding. By integrating wavelength division multiplexing technology of a virtual imaging phased array with spatial spectroscopy of a grating, broadband spectra of an ultrafast laser pulse are mapped onto a 2D plane. Subsequently, Galilean structure is employed for beam expansion, thereby achieving a large field-of-view 2D imaging capability. A dispersion Fourier transform module is also incorporated into the system, enabling the mapping of wavelength to time, facilitating ultrafast imaging through the detection of pulse signals by high-speed photodetector. The experimental setup employs a home-made mode-locked laser with a spectral bandwidth of 12 nm, capable of large-field imaging of 6 × 20 mm², representing a 20-fold expansion compared to the unexpanded field of view, and an imaging frame rate of up to 7.75 MHz is obtained. The dynamic imaging capability of the system is demonstrated through imaging a freely falling sphere. The proposed imaging system, with its large field of view and high frame rate, is ideal for high-speed flow diagnostics.