Hussain (Md. Abu) Nyeem received a B.Sc. degree in electronics and communication engineering (ECE) from the Khulna University of Engineering & Technology (KUET), Bangladesh in 2007 and a Ph.D. degree in electrical engineering and computer science (EECS) from the Queensland University of Technology (QUT), Australia, in 2014 with high distinction and nomination for the QUT best Ph.D. thesis award.
Dr. Nyeem is an Associate Professor at the Military Institute of Science and Technology (MIST), Bangladesh. Before joining MIST in 2015, he held positions of Assistant Professor and Lecturer at KUET, Bangladesh; Sessional Academic and Doctoral Research Fellow at QUT, Australia; Guest Lecturer at Khulna University and Bangladesh University of Professionals; and Exchange Research Scholar at the University of Fukui, Japan in his fifteen years of profession.
His research interest broadly revolves around image processing and its applications. Mainly, his work concentrates on the processing of the spatial and frequency domain image-features, including their extraction and classification, and also on the biomedical image segmentation and analysis. Development of the intelligent and efficient methods for detection of road-lane, vehicle licence-plate, lip-movement, and representation of objects like the segmented region-of-interest (RoI) with polygonal decimation of contours have been endeavoured. Besides, new visual information protection techniques, including several data hiding schemes and their multimedia applications, are the outcome of his recent work. Multidisciplinary research collaboration has also extended the scope of his work to the computational analysis and visual representation of extensive environmental (i.e., water, soil, air and food-plants) data for heavy-metals’ assessment. He has further interest in developing the intelligent, bio-inspired and vision-based decision support system.
Dr. Nyeem has been leading the Visual Information Processing research-group at MIST since 2015 and has supervised eight masters (four ongoing) and more than twenty-five honours (final-year) projects and theses. His research projects have been funded by national universities and international postgraduate and exchange research programs. He has published a total of 43 peer-reviewed articles since 2012 in reputable journals, book series and proceedings. His recent (in the last five years) research endeavour has been recognized in the leading international conferences (including the 21st IEEE ICCIT 2018 and IEEE Region 10 Symposium 2020) with five best-paper and one best-presentation awards.
Dr. Nyeem is currently an Academic Editor of PLOS ONE, Senior Member of the IEEE and Fellow of the Institute of Engineers, Bangladesh (IEB). He has been actively contributing to the national academic and research initiatives. He chaired technical sessions and was on the technical program committee of more than 15 international conferences, including the recent IC4IR’21, ICEEICT’21, and IEEE ICTP’21. Likewise, he also intermittently reviews articles of the leading journals, including IEEE Transactions on Systems, Man, and Cybernetics: Systems, IEEE Access, and IET Image Processing, and received the Elsevier outstanding reviewer (Optik) award in 2018.
Interests
Image Processing
Image Analysis
Data Hiding
Medical AI
eHealth
Education
PhD in Electrical, Electronic and Communication Engineering
— Queensland University of Technology, Australia
(2010-2014)
B.Sc. in Electronics and Communication Engineering
— Khulna University of Engineering and Technology, Bangladesh
(2001-2007)
Selected Publications
Utilizing Reverse Attention for Enhanced Mitochondria Segmentation in Microscopic Images2025
Mitochondrial segmentation in microscopic images is a complex task hindered by variations in mitochondrial size, color, texture, and their often ambiguous borders with surrounding tissue. To overcome these obstacles, we adapted Parallel Reverse Attention Network (Pranet), originally developed for precise polyp detection, to the domain of mitochondrial segmentation. Pranet employs Parallel Partial Decoder (PPD) to integrate multi-level features, producing a high-level context map that guides subsequent processing. The model further enhances segmentation precision through a Reverse Attention (RA) module, which adeptly identifies object boundaries by iteratively refining region-boundary interactions. Our approach was thoroughly evaluated on the EPFL dataset, yielding substantial improvements in segmentation accuracy. Additionally, Pranet demonstrated robust generalization and real-time performance, underscoring its suitability for comprehensive microscopic image analysis.
BibTeX
Click to copy
@INPROCEEDINGS{11013970,
author={Alam, Ayat Subah and Rhythm, Faiaz Hasanuzzaman and Ovi, Tareque Bashar and Bashree, Nomaiya and Nyeem, Hussain and Wahed, Md Abdul},
booktitle={2025 International Conference on Electrical, Computer and Communication Engineering (ECCE)},
title={Utilizing Reverse Attention for Enhanced Mitochondria Segmentation in Microscopic Images},
year={2025},
volume={},
number={},
pages={1-6},
keywords={Image segmentation;Adaptation models;Mitochondria;Accuracy;Image color analysis;Microscopy;Computer architecture;Transformers;Real-time systems;Biomedical imaging;deep learning;medical image segmentation;contextual information;mitochondria segmentation;electron-microscopy images;reverse attention},
doi={10.1109/ECCE64574.2025.11013970}}
Optimizing Monocular Depth Estimation through Bi-Level Nested Architecture Integration2025
Monocular depth estimation (MDE) has evolved into a vital research focus in visual computing and autonomous robotic systems. While recent deep learning approaches have advanced the field, existing methods often struggle to jointly incorporate global context and fine-grained local details, and may be hampered by complex, computationally expensive network architectures. To address these challenges, we employ the U2Net framework, which incorporates Residual U-Blocks designed to adapt the area of reception and capture multi-scale contextual data. By nesting U-shaped structures, U2Net effectively increases feature extraction depth while maintaining computational efficiency, thus offering a more balanced and scalable solution for MDE. Our proposed model, developed using the NYUDepth V2 dataset, exceeds contemporary methods in critical assessment measures, including RMSE, RMSE(log), and Absolute Relative Error (Abs Rel). These improved results highlight the robustness and precision of proposed model or a range of practical applications in robotics, augmented reality, and other domains requiring real-time scene geometry understanding.
BibTeX
Click to copy
@INPROCEEDINGS{11013114,
author={Rhythm, Faiaz Hasanuzzaman and Bashar Ovi, Tareque and Bashree, Nomaiya and Islam Ratul, Md. Raisul and Nyeem, Hussain and Wahed, Md Abdul},
booktitle={2025 International Conference on Electrical, Computer and Communication Engineering (ECCE)},
title={Optimizing Monocular Depth Estimation through Bi-Level Nested Architecture Integration},
year={2025},
volume={},
number={},
pages={1-6},
keywords={Deep learning;Adaptation models;Visualization;Depth measurement;Computational modeling;Reliability engineering;Robustness;Real-time systems;Computational efficiency;Robots;deep learning;depth estimation;contextual and local information;residual u-block;nested u-structure},
doi={10.1109/ECCE64574.2025.11013114}}
Bridging Classical and Quantum Models via Attention-Guided Feature Distillation2025
While Quantum Neural Networks (QNNs) offer the potential for resource-efficient image classification, practical implementations are hindered by qubit instability and finite availability, causing them to often underperform classical counterparts. This performance disparity constitutes a critical research gap. We propose a novel hybrid classical-quantum knowledge distillation approach to bridge this gap. Our method utilizes a pre-trained classical CNN (AlexNet) as a teacher to guide the training of a quantum student model by minimizing the Kullback-Leibler (KL) divergence (DKL) between their output distributions. The quantum student model employs an architecture designed to facilitate intricate feature interactions using iterated CNOT gates. This technique substantially improves quantum model accuracy, demonstrating gains of 0.34% (6-class, 3 qubits) and 0.28% (8-class, 3 qubits) on the MNIST dataset, and 0.74% (8-class, 4 qubits). Our contribution lies in operationalizing a knowledge distillation paradigm for applied Quantum Machine Learning (QML), enabling the effective transfer of robust knowledge from established classical architectures to resource-constrained quantum models, thereby enhancing faster convergence and accelerating practical QML deployment.
BibTeX
Click to copy
@INPROCEEDINGS{11159974,
author={Ovi, Tareque Bashar and Bashree, Nomaiya and Rhythm, Faiaz Hasanuzzaman and Chowdhury, Disha and Nyeem, Hussain and Wahed, Md Abdul},
booktitle={2025 2nd International Conference on Next-Generation Computing, IoT and Machine Learning (NCIM)},
title={Bridging Classical and Quantum Models via Attention-Guided Feature Distillation},
year={2025},
volume={},
number={},
pages={1-6},
keywords={Training;Accuracy;Qubit;Transfer learning;Neural networks;Computer architecture;Logic gates;Quantum networks;Next generation networking;Image classification;Deep learning;knowledge distillation;amplitude encoding;transfer learning;QML},
doi={10.1109/NCIM65934.2025.11159974}}
No images
YESnet: YOLOv11 Enabled SAM-2 Framework for Memory-Efficient Skin Lesion Segmentation2025
This paper presents YESnet, a novel framework integrating the YOLOv11 object detector with the Segment Anything Model 2 (SAM-2) for memory-efficient skin lesion segmentation. The proposed architecture introduces: (1) An anchor-free YOLOv11 detection head that eliminates predefined bounding box priors while maintaining detection accuracy, (2) A memory-optimized SAM-2 variant that preserves boundary delineation capabilities through compressed feature retention, and (3) A lightweight co-processing pipeline that coordinates detection and segmentation stages without computational redundancy. Evaluated on the HAM10000 dataset, YESnet achieves competitive performance in lesion segmentation (96.00% accuracy, 92.38% Dice) while demonstrating efficient memory utilization. The framework's design enables effective lesion boundary capture with minimal parameter overhead, making it particularly suitable for resource-constrained clinical deployment scenarios.
BibTeX
Click to copy
@INPROCEEDINGS{11171738,
author={Bashree, Nomaiya and Ovi, Tareque Bashar and Zahid, Sadia Binte and Rhythm, Faiaz Hasanuzzaman and Nyeem, Hussain and Abdul, Md},
booktitle={2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)},
title={YESnet: YOLOv11 Enabled SAM-2 Framework for Memory-Efficient Skin Lesion Segmentation},
year={2025},
volume={},
number={},
pages={1-5},
keywords={Image segmentation;Accuracy;Image analysis;Pipelines;Redundancy;Memory management;Detectors;Skin;Lesions;Photonics;Deep learning;skin lesion segmentation;foundational model;nano YOLOv11 architecture;SAM-2},
doi={10.1109/QPAIN66474.2025.11171738}}
Enhancing U2Net for Precise Road Extraction from Satellite Images via Channel Refinement2025
Road extraction from very high-resolution satellite imagery (VHR) is necessary for remote sensing applications like city management, and navigation. In this study, we introduce SE-U2Net, a novel deep learning model for precise road extraction. Our proposed model integrates the Squeeze and Excitation (SE) network with the U2 Net architecture for enhanced selective channel refinement and better contextual information capture. SE-U2Net employs Residual U-blocks (RSU), which combine receptive fields of different sizes, enabling the model to capture multiscale contextual data efficiently. The baseline U2Net has 75.06 GFLOPs and 37.51 GMacs, while SE-U2Net maintains this efficiency with 75.07 GFLOPs and 37.51 GMacs, despite its enhanced representational capacity through SE blocks. Emperical outcomes on the DeepGlobe dataset demonstrate that SE-U2Net outperforms the base U2Net model and other state-of-the-art models, achieving superior performance in terms of Dice coefficient, mean Intersection over Union (mIoU), precision, and recall. These results underscore SE-U2Net's potential as a robust tool for high-accuracy road segmentation from VHR satellite imagery.
BibTeX
Click to copy
@INPROCEEDINGS{11171628,
author={Rhythm, Faiaz Hasanuzzaman and Ovi, Tareque Bashar and Bashree, Nomaiya and Nyeem, Hussain and Wahed, Md Abdul},
booktitle={2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)},
title={Enhancing U2Net for Precise Road Extraction from Satellite Images via Channel Refinement},
year={2025},
volume={},
number={},
pages={1-6},
keywords={Deep learning;Image segmentation;Adaptation models;Roads;Computational modeling;Urban areas;Computer architecture;Satellite images;Computational efficiency;Context modeling;Road extraction;Segmentation;Skip connection;Squeeze-excitation;High-resolution satellite image},
doi={10.1109/QPAIN66474.2025.11171628}}
DoubleUNet++: Channel-Aware Gated Attention for Road Extraction in Satellite Imagery2025
Road extraction from satellite imagery faces critical challenges in capturing contextual spatial relationships and mitigating background interference. We introduce DoubleUNet++, which improves the DoubleUNet architecture with two main enhancements: (1) A channel-aware gated attention (CAGA) mechanism in the skip connections of the second U-Net, dynamically weighting channel features via decoder gating signals, and (2) Squeeze-and-Excitation layers in the attention path for adaptive channel-wise feature recalibration. Evaluated on the CHN6-CUG dataset (4,511 annotated images), our model achieves state-of-the-art performance with 78.52 Dice score and 64.94 mIoU - representing 3.32 % and 3.74 % absolute gains over baseline DoubleUNet. Despite these improvements, DoubleUNet++ maintains similar computational complexity (109.33 GFLOPs, 54.64 GMacs) compared to the baseline (107.41 GFLOPs, 53.68 GMacs). The architecture maintains comparable complexity (29.5M parameters) while significantly improving precision (81.45% vs 73.78%) through refined spatial focus on roadrelevant regions. Qualitative analysis demonstrates the CAGA mechanism's effectiveness in reducing misclassification from nonroad objects such as vegetation and buildings while preserving road topology. These advancements establish DoubleUNet++ as a robust solution for large-scale road network mapping without requiring additional annotated data.
BibTeX
Click to copy
@INPROCEEDINGS{11172088,
author={Rhythm, Faiaz Hasanuzzaman and Bashree, Nomaiya and Ovi, Tareque Bashar and Nyeem, Hussain and Wahed, Md Abdul},
booktitle={2025 International Conference on Quantum Photonics, Artificial Intelligence, and Networking (QPAIN)},
title={DoubleUNet++: Channel-Aware Gated Attention for Road Extraction in Satellite Imagery},
year={2025},
volume={},
number={},
pages={1-6},
keywords={Image segmentation;Roads;Vegetation mapping;Computer architecture;Logic gates;Feature extraction;Data models;Spatial databases;Satellite images;Topology;Deep learning;Road extraction;Cascaded model;Segmentation;Gated attention;Skip connection;Squeeze-excitation;High-resolution satellite image},
doi={10.1109/QPAIN66474.2025.11172088}}
A Novel Dual Attention Approach for DNN Based Automated Diabetic Retinopathy Grading2024
Diabetic retinopathy (DR) poses a serious threat to vision, emphasising the need for early detection. Manual analysis of fundus images, though common, is error-prone and time-intensive. Existing automated diagnostic methods lack precision, particularly in the early stages of DR. This paper introduces the Soft Convolutional Block Attention Module-based Network (Soft-CBAMNet), a deep learning network designed for severity detection, which features Soft-CBAM attention to capture complex features from fundus images. The proposed network integrates both the convolutional block attention module (CBAM) and the soft-attention components, ensuring simultaneous processing of input features. Following this, attention maps undergo a max-pooling operation, and refined features are concatenated before passing through a dropout layer with a dropout rate of 50%. Experimental results on the APTOS dataset demonstrate the superior performance of Soft-CBAMNet, achieving an accuracy of 85.4% in multiclass DR grading. The proposed architecture has shown strong robustness and general feature learning capability, achieving a mean AUC of 0.81 on the IDRID dataset. Soft-CBAMNet's dynamic feature extraction capability across all classes is further justified by the inspection of intermediate feature maps. The model excels in identifying all stages of DR with increased precision, surpassing contemporary approaches. Soft-CBAMNet presents a significant advancement in DR diagnosis, offering improved accuracy and efficiency for timely intervention.
BibTeX
Click to copy
@article{https://doi.org/10.1002/ima.23175,
author = {Ovi, Tareque Bashar and Bashree, Nomaiya and Nyeem, Hussain and Wahed, Md Abdul and Rhythm, Faiaz Hasanuzzaman and Alam, Ayat Subah},
title = {A Novel Dual Attention Approach for DNN Based Automated Diabetic Retinopathy Grading},
journal = {International Journal of Imaging Systems and Technology},
volume = {34},
number = {5},
pages = {e23175},
doi = {https://doi.org/10.1002/ima.23175},
url = {https://onlinelibrary.wiley.com/doi/abs/10.1002/ima.23175},
eprint = {https://onlinelibrary.wiley.com/doi/pdf/10.1002/ima.23175},
note = {e23175 IMA-24-504.R1},
abstract = {ABSTRACT Diabetic retinopathy (DR) poses a serious threat to vision, emphasising the need for early detection. Manual analysis of fundus images, though common, is error-prone and time-intensive. Existing automated diagnostic methods lack precision, particularly in the early stages of DR. This paper introduces the Soft Convolutional Block Attention Module-based Network (Soft-CBAMNet), a deep learning network designed for severity detection, which features Soft-CBAM attention to capture complex features from fundus images. The proposed network integrates both the convolutional block attention module (CBAM) and the soft-attention components, ensuring simultaneous processing of input features. Following this, attention maps undergo a max-pooling operation, and refined features are concatenated before passing through a dropout layer with a dropout rate of 50\%. Experimental results on the APTOS dataset demonstrate the superior performance of Soft-CBAMNet, achieving an accuracy of 85.4\% in multiclass DR grading. The proposed architecture has shown strong robustness and general feature learning capability, achieving a mean AUC of 0.81 on the IDRID dataset. Soft-CBAMNet's dynamic feature extraction capability across all classes is further justified by the inspection of intermediate feature maps. The model excels in identifying all stages of DR with increased precision, surpassing contemporary approaches. Soft-CBAMNet presents a significant advancement in DR diagnosis, offering improved accuracy and efficiency for timely intervention.},
year = {2024}
}
A Transformer-based Text-Guided Approach for Improved Colonoscopic Polyp Segmentation2024
Automated colonoscopic polyp segmentation, vital for early colorectal cancer detection, has notably advanced with deep learning. However, reliance on operator expertise and challenges posed by varied polyp sizes during training, along with limitations in representing long-range dependencies, warrant further refinement in the field. In this paper, we propose a novel transformer-based text guided architecture called TGA-TransNet to address these challenges. Particularly, the TGA-TransNet model incorporates size-related and polyp number-related information through text attention during training. Additionally, Transformer layers are utilized to extract global contexts. Our early experimental results on the popular Kvasir-SEG dataset indicate that the proposed integration of text embeddings and Transformer enhances the overall segmentation performance of the model, surpassing its state-of-the-art counter-part. Feature map analysis also suggests that our proposed TGA-TransNet demonstrates strong generalization capabilities across the variable-sized polyps.
BibTeX
Click to copy
@INPROCEEDINGS{10561827,
author={Ovi, Tareque Bashar and Bashree, Nomaiya and Nyeem, Hussain and Wahed, Md Abdul and Alam, Ayat Subah and Rhythm, Faiaz Hasanuzzaman},
booktitle={2024 3rd International Conference on Advancement in Electrical and Electronic Engineering (ICAEEE)},
title={A Transformer-based Text-Guided Approach for Improved Colonoscopic Polyp Segmentation},
year={2024},
volume={},
number={},
pages={1-6},
keywords={Training;Deep learning;Image segmentation;Image coding;Colonic polyps;Predictive models;Transformers;deep learning;medical image segmentation;contextual information;text guided attention;transformer},
doi={10.1109/ICAEEE62219.2024.10561827}}
🧪 Ongoing Research Projects
Investigating Adversarial Resilience of Popular Models For Polyp And Skin Lesion Segmentation
The 5th International Conference on Trends in Electronics and Health Informatics (TEHI 2025)
, 2025
Objective
This study evaluates the robustness of leading segmentation models against advanced adversarial attacks in the context of polyp and skin lesion identification.
Key Findings
We benchmark the robustness of leading segmentation models under strong adversarial attacks, systematically identifying their failure modes in the context of polyp and skin lesion tasks.
The first systematic study of bi-level nested architectures under adversarial threat, clarifying how multiscale fusion and cross-level aggregation influence resilience.
Evidence that global feature representations mitigate adversarial perturbations, with distilled design guidelines for more robust medical segmentation systems.
Photos
Comparative Insight of Scalable Graph Based Spatial Domain Discovery From Stereo-Seq and Slide-Seq Data
Taylor and Francis Book Chapter, 3rd International Conference on Big Data, IoT and Machine Learning (BIM 2025)
, 2025
Objective
To systematically evaluate the performance and scalability of leading GNN-based spatial transcriptomics (ST) models—STAGATE, GraphST, and SEDR—on high-density next-generation datasets such as Stereo-seq and Slide-seq, and to identify their limitations in modeling complex spatial structures.
Key Findings
Performance Evaluation: Comprehensive assessment using ARI, AMI, NMI, and HOMO metrics revealed modest and variable clustering performance across high-density datasets.
- SEDR achieved the highest scores on Slide-seq datasets, but overall accuracy remained low.
- Stereo-seq results were inconsistent, with no single model consistently outperforming others.
Scalability Limitation: Existing GNN-based ST models struggle to generalize to high-resolution data, indicating poor scalability to large, complex tissue structures.
Benchmark Establishment: This study provides a performance baseline for evaluating future models on next-generation ST platforms.
Research Implication: Highlights the urgent need for developing new computational frameworks that are more robust, adaptive, and scalable for analyzing dense spatial transcriptomics data.
Photos
Ensemble Based Graph Attention Auto-Encoder Architecture For Unsupervised Spatial Clustering
Taylor and Francis Book Chapter, 3rd International Conference on Big Data, IoT and Machine Learning (BIM 2025)
, 2025
Objective
To overcome the limitation of fixed spatial neighborhood scales in existing spatial transcriptomics (ST) models by developing En-STAGATE, an ensemble framework that integrates multiple spatial graphs across varying neighborhood radii to capture both local and global tissue structures for improved domain identification.
Key Findings
Multi-Scale Ensemble Design: En-STAGATE constructs and fuses representations from multiple spatial graphs at different neighborhood radii, effectively modeling both fine-grained cellular interactions and large-scale tissue organization.
Enhanced Biological Interpretability: The multi-scale integration provides a more comprehensive and biologically meaningful latent representation of tissues.
Empirical Validation: Achieved competitive or superior performance across 14 benchmark datasets, outperforming STAGATE and other leading methods in several key cases.
Improved Domain Delineation: Notably higher Adjusted Rand Index (ARI) scores on datasets such as 151674, 151675, 151676, and Mouse, demonstrating robustness in complex tissue structures.
Key Insight: Highlights that a single neighborhood scale is insufficient for spatial domain identification, establishing multi-scale analysis as a critical paradigm in spatial transcriptomics.
Photos
Proximity Enhanced Multi-Modal Graph Based Spatial Transcriptomics Clustering
Taylor and Francis Book Chapter, 3rd International Conference on Big Data, IoT and Machine Learning (BIM 2025)
, 2025
Objective
To address the over-smoothing and spatial bias in current spatial transcriptomics (ST) models by developing GraphST++, a multi-modal framework that integrates spatial, gene expression, and histological features for constructing biologically faithful tissue graphs.
Key Findings
Multi-Modal Graph Construction: GraphST++ defines inter-spot relationships using a composite score combining spatial distance, gene expression similarity, and histological morphology, reducing over-reliance on proximity-based connections.
Improved Tissue Representation: The proposed approach captures complex biological interactions and produces a more accurate reflection of the tissue microenvironment.
Superior Performance: Demonstrated higher Adjusted Rand Index (ARI) compared to leading models, including GraphST, across multiple benchmark datasets such as human breast cancer (BRCA) and human brain tissues.
Robustness Across Datasets: While not universally dominant, GraphST++ consistently benefits from morphological integration, yielding more stable and biologically interpretable clustering results.
Broader Implications: Highlights the importance of multi-modal fusion in advancing spatial transcriptomics analysis and enabling more precise tissue domain identification.
Photos
Performance Analysis of Semi-Supervised Frameworks for Polyp Segmentation
The 5th International Conference on Trends in Electronics and Health Informatics (TEHI 2025)
, 2025
Objective
In this study, we systematically investigate how established supervised segmentation models perform within three representative semi-supervised frameworks: Min-Max Similarity, Duo-SegNet, and Deep Co-Training.
Key Findings
We represent a comprehensive evaluation of multiple supervised segmentation architectures in semi-supervised learning scenarios.
We identify model characteristics that enable effective utilization of unlabeled data within semi-supervised frameworks.
We provide empirical insights that inform the design of more better and dependable semi-supervised polyp segmentation methods, thereby reducing annotation requirements while maintaining high segmentation accuracy.
Photos
UAPNet: Uncertainty Augmented Pyramid Vision Transformer Network With Efficient Channel Filtering for Polyp Segmentation
The 5th International Conference on Trends in Electronics and Health Informatics (TEHI 2025)
, 2025
Objective
To mitigate feature redundancy in deep networks, we integrate the Efficient Channel Attention (ECA) module with the PVT encoder and augment it with an Uncertainty Augmented Context Attention (UACA) mechanism.
Key Findings
PVT with Uncertainty-Augmented Context Attention. We integrate the PVT encoder with an uncertainty-augmented context attention mechanism, enabling the model to focus explicitly on ambiguous boundary regions. This mechanism improves the accuracy of boundary delineation by using uncertainty as a guiding signal.
Integrating ECA with PVT Based Encoder. We incorporate ECA after each PVT stage, which refines feature representations while maintaining the original dimensionality. This enhancement improves the model’s discriminative power without introducing additional computational overhead.
Explainability for Clinical Reliability. To ensure transparency and clinical applicability, we apply explainable AI (XAI) techniques, including uncertainty maps and heatmaps for foreground and edge regions. These methods improve interpretability and provide clinicians with more reliable decision making tools by highlighting areas of uncertainty and focus during polyp segmentation.
Photos
Transformer Enhanced Graph Based Spatial Domain Identification
Taylor and Francis Book Chapter, 3rd International Conference on Big Data, IoT and Machine Learning (BIM 2025)
, 2025
Objective
To overcome the limitation of existing graph-based spatial transcriptomics (ST) models in capturing long-range spatial dependencies by developing STAGATE++, a hybrid GAT–Transformer architecture that jointly learns local and global tissue representations.
Key Findings
Hybrid Architecture Advantage: STAGATE++ integrates a Graph Attention Network (GAT) for precise local feature extraction with a Transformer encoder for modeling long-range spatial dependencies.
- Enhanced Spatial Awareness: The model constructs a holistic and spatially coherent tissue representation, improving understanding of tissue organization.
Empirical Superiority: Demonstrated state-of-the-art performance across 14 benchmark datasets, consistently outperforming existing methods in spatial domain identification.
Validation Metric: Significant improvements in Adjusted Rand Index (ARI) confirm the importance of global context modeling in tissue delineation.
Future Extensions: Plans include incorporating histopathological image features, applying self-supervised training for better generalization, and integrating GNN interpretability techniques for biological insight and clinical transparency.
