research

2025

1. 

   Enhancing Self-Supervised Learning for Image Clustering Using Geometric Deep Learning

   Md Kamrul Islam, Xianyun Zhuang, Akash Malhotra, and 1 more author

   Zenodo, 2025

   Abs DOI Bib PDF Code

   Contemporary self-supervised learning approaches have significantly advanced unsupervised image analysis, yet their application to medical imaging remains constrained by conventional CNN architectures’ inherent limitations in handling geometric transformations. This research introduces a novel deep clustering framework that incorporates Group Equivariant Convolutional Neural Networks (G-CNNs) to substantially enhance feature representation learning for medical image clustering. Our methodological framework implements a systematic iterative pipeline extending the DeepCluster architecture through dual feature extraction pathways: (1) an AlexNet implementation incorporating Sobel filtering with random rotation augmentation, and (2) a P4M-equivariant CNN architecture that inherently encodes rotation and reflection invariance properties. Evaluated on the NIH Chest X-ray dataset, our approach leverages k-means clustering with iterative pseudo-labeling to progressively refine latent representations. Experimental results demonstrate that the proposed G-CNN framework achieves improved semantic coherence and clustering quality, converging faster than the conventional CNN-based approach. These findings establish the significant potential of geometric deep learning techniques for developing more robust and clinically relevant solutions for medical image clustering and automated disease detection systems.

   @article{kamrul2025gcnn,
     title = {Enhancing Self-Supervised Learning for Image Clustering Using Geometric Deep Learning},
     author = {Islam, Md Kamrul and Zhuang, Xianyun and Malhotra, Akash and {Seghouani Bennacer}, Nac\'{e}ra},
     year = {2025},
     publisher = {Zenodo},
     journal = {Zenodo},
     doi = {10.5281/zenodo.17494110},
     url = {https://doi.org/10.5281/zenodo.17494110},
   }

2024

1. 

   Retrieval Augmented Generation for Knowledge Graphs

   Md Kamrul Islam and David García Morillo

   Zenodo, 2024

   Abs DOI Bib PDF

   The field of Knowledge Graphs (KGs) has seen significant advancements in recent years, driven by the need to structure and query large-scale, complex datasets effectively. In parallel, another promising technique has emerged in the realm of NLP and Large Language Models (LLM), Retrieval Augmented Generation (RAG). These leverage the power of non-parametric external knowledge bases to enhance the reliability, accuracy, and completeness of LLMs. This literature review aims to provide a comprehensive overview of the state-of-the-art methodologies and applications of RAG, with a focus in the context of KGs. We explore various approaches to integrating retrieval mechanisms with generative processes, highlighting how these hybrid models address challenges such as knowledge incompleteness, real-time data updates, and natural language understanding. Furthermore, we examine the efficacy of RAG techniques in different domains, including finance, healthcare and education, among others, illustrating their potential to exploit the semantics and richness of KGs. Key research contributions, ongoing challenges, and future directions in the deployment of RAG for KGs are also discussed, providing a roadmap for researchers and practitioners seeking to leverage these technologies in their work.

   @article{kamrul2024ragkg,
     title = {Retrieval Augmented Generation for Knowledge Graphs},
     author = {Islam, Md Kamrul and {Garc\'{i}a Morillo}, David},
     year = {2024},
     publisher = {Zenodo},
     journal = {Zenodo},
     doi = {10.5281/zenodo.17494287},
     url = {https://doi.org/10.5281/zenodo.17494287},
   }

2021

1. 

   Brain Tumor Detection and Classification using CNN

   Md Kamrul Islam

   2021

   Abs DOI Bib PDF Code

   This paper presents a new CNN architecture for brain tumor classification of three types: meningioma, glioma, pituitary tumor, and no tumor from T1-weighted contrast-enhanced magnetic resonance images. The experimental result shows that the proposed CNN model achieves 98% validation accuracy, outperforming VGG-16 (93%), Xception (95%), ResNet-50 (94%), and Inception-V3 (97%). The proposed CNN model requires significantly less computational power and has much better accuracy than other pre-trained models.

   @thesis{kamrul2021braintumor,
     title = {Brain Tumor Detection and Classification using {CNN}},
     author = {Islam, Md Kamrul},
     year = {2021},
     publisher = {Zenodo},
     school = {Sichuan University},
     type = {Undergraduate Thesis},
     doi = {10.5281/zenodo.17494372},
     url = {https://doi.org/10.5281/zenodo.17494372},
   }