ED-Net: A Hybrid Deep Learning Framework for Precise Polyp Segmentation Integrating EfficientNetB0 with Double U-Net Architecture and Atrous Spatial Pyramid Pooling with Test-Time Augmentation
Rahul Kumar Kundu, Hafsa Binte Kibria, Md Faysal Ahamed, and 1 more author
Colorectal cancer (CRC) is a major global health concern, and colorectal polyps are early indicators of CRC. The precise and timely identification of these polyps is crucial for reducing CRC-related mortality rates. This study introduces a deep-learning-based segmentation framework, ED-Net, that incorporates a pretrained EfficientNetB0 encoder within a dual encoder-decoder architecture. This architecture is further enhanced by Atrous Spatial Pyramid Pooling (ASPP) for multi-scale feature extraction and squeeze-and-excitation (SE) blocks for adaptive channel-wise attention. Additionally, Test Time Augmentation (TTA) with horizontal and vertical flips was applied, which improved the interior and boundary segmentation accuracy through ensemble averaging. ED-Net was trained and tested on the Kvasir-SEG dataset, and to evaluate its generalization ability, cross-dataset evaluations were conducted on three benchmark datasets: CVC-ClinicDB, ETIS-LaribPolypDB, and PolypGen2021. ED-Net with TTA achieved an accuracy of 97.13%, a Dice Score of 92.26%, an IoU of 91.08%, a Frame per Second (FPS) of 12.93 with TTA and 47.17 without TTA on Kvasir-SEG, substantially outperforming the SOTA models in terms of accuracy and computational complexity. These results demonstrate that the model is effective, robust, and has potential for clinical deployment for the early detection of CRC.
@article{kundu2026ednet,title={ED-Net: A Hybrid Deep Learning Framework for Precise Polyp Segmentation Integrating {EfficientNetB0} with Double {U-Net} Architecture and Atrous Spatial Pyramid Pooling with Test-Time Augmentation},author={Kundu, Rahul Kumar and Kibria, Hafsa Binte and Ahamed, Md Faysal and Chowdhury, Muhammad EH},journal={Biomedical Signal Processing and Control},volume={123},pages={110566},year={2026},issn={1746-8094},publisher={Elsevier},doi={10.1016/j.bspc.2026.110566},keyword={journal}}
Conference Proceedings
2025
IEEE STI
Beyond the Leaf: Intelligent Lightweight Disease Detection in Tea Plants with Spatial Convolutional Neural Networks and Explainable AI
Reazul Hasan Prince, Md. Omaer Faruq Goni, Hasibul Islam Peyal, and 1 more author
In 2025 IEEE 7th International Conference on Sustainable Technologies For Industry 5.0 (STI), 2025
Tea is one of the classic beverages in Bangladesh. Since its demand and economic contribution are significant, agronomists need to quickly identify tea diseases and address them. Traditional methods for diagnosing these diseases depend on pathologists to identify symptoms visually, which often results in misdiagnosis and lengthy detection processes. A Convolutional Neural Network (CNN) has been proposed to address this. The proposed CNN model successfully classifies six types of tea leaf diseases with 99.23% accuracy, outperforming MobileNetV2 (91.23%) and Xception (91.66%) on the same dataset. The model achieves high classification performance with a significantly reduced number of parameters, resulting in a lightweight architecture. Model performance was evaluated using precision, recall, F1 score, confusion matrix, ROC, PR curve, and MCC value. Grad-CAM visualization was applied to explain the model’s predictions.
@inproceedings{11367552,title={Beyond the Leaf: Intelligent Lightweight Disease Detection in Tea Plants with Spatial Convolutional Neural Networks and Explainable {AI}},author={Prince, Reazul Hasan and Goni, Md. Omaer Faruq and Peyal, Hasibul Islam and Kundu, Rahul Kumar},booktitle={2025 IEEE 7th International Conference on Sustainable Technologies For Industry 5.0 (STI)},publisher={IEEE},year={2025},pages={1--6},doi={10.1109/STI69347.2025.11367552},keyword={conference}}
ICCIT
Advancing Cybersecurity: A Study on Phishing Website Detection Using Machine Learning
Syed Nazmus Sakib, Oishi Jyoti, and Rahul Kumar Kundu
In 2025 28th International Conference on Computer and Information Technology (ICCIT), 2025
Phishing attacks remain one of the most prevalent cybersecurity threats, often leading to financial loss and identity theft. Traditional detection mechanisms, such as blacklists or heuristic-based filters, are insufficient against zero-day phishing URLs and adversarial manipulation. This paper proposes a robust ensemble-based phishing website detection system that integrates XGBoost, LightGBM, CatBoost, and TabPFN, focusing exclusively on URL-based lexical and structural features. A Kaggle dataset of 11,055 samples was used, partitioned into 8,955 training and 2,100 testing instances. Preprocessing involved SelectKBest for feature selection, StandardScaler for normalization, and SMOTE for class balancing. Experimental evaluation shows the proposed ensemble achieves 93.3% accuracy and 0.99 recall, outperforming individual models while ensuring interpretability through SHAP. Results demonstrate that combining boosting techniques with modern tabular deep learning improves generalization and reliability, making the model suitable for real-time phishing detection.
@inproceedings{11491653,title={Advancing Cybersecurity: A Study on Phishing Website Detection Using Machine Learning},author={Sakib, Syed Nazmus and Jyoti, Oishi and Kundu, Rahul Kumar},booktitle={2025 28th International Conference on Computer and Information Technology (ICCIT)},publisher={IEEE},year={2025},pages={3530--3535},doi={10.1109/ICCIT68739.2025.11491653},keyword={conference}}
Under Review
2026
IEEE OJCS
Demographic Fairness in Skin Lesion Segmentation: Lesion-Type Confounding of Age-Related Performance Disparities
Rahul Kumar Kundu, Md Masum Mia, Amith Khandakar, and 3 more authors
2026
Under Review — IEEE Open Journal of the Computer Society
@unpublished{kundu2025metaseg,title={Demographic Fairness in Skin Lesion Segmentation: Lesion-Type Confounding of Age-Related Performance Disparities},author={Kundu, Rahul Kumar and Mia, Md Masum and Khandakar, Amith and Shi, Zumin and Kibria, Hafsa Binte and Ayari, Mohamed Arselene},year={2026},note={Under Review — IEEE Open Journal of the Computer Society},keyword={progress}}
CMPB
Data Leakage in Clinical Machine Learning: A Systematic Review of Sources, Impacts and Mitigation Strategies
@unpublished{kundu2026dataleakage,title={Data Leakage in Clinical Machine Learning: A Systematic Review of Sources, Impacts and Mitigation Strategies},author={Kibria, Nafisa Binte Ghulam and Kibria, Hafsa Binte and Kundu, Rahul Kumar},year={2026},note={Under Review — Computer Methods and Programs in Biomedicine, Elsevier (Q1)},keyword={progress}}