Selected peer-reviewed publications, preprints, and manuscripts under review are listed below. To view the complete list of publications, please visit my Google Scholar profile.
Authors: Shadman Sobhan, Kazi Abrar Mahmud, Abduz Zami, Mohiuddin Ahmed, Md Palash Uddin
Journal: Biomedical Signal Processing and Control Q1 IF 5.7
MedPrompt combines a few-shot-prompted Large Language Model (LLM) with a modular Convolutional Neural Network (DeepFusionLab). The user provides an image along with a text prompt describing the desired task. Based on the prompt, the LLM determines whether the CNN should operate in classification or segmentation mode and selects the corresponding task-specific model weights. This process of selecting the appropriate model weights is referred to as “Weight Routing.” This architecture offers several advantages. Since task-specific weights are trained separately and loaded according to the requested task, new classification or segmentation tasks can be incorporated without retraining the entire model from scratch. Furthermore, using task-specific weights allows the model to be optimized for each individual task, potentially providing better performance than using a single set of general-purpose weights.
Authors: Abduz Zami, Shadman Sobhan, Mohiuddin Ahmed, Md Palash Uddin
Journal: Biomedical Signal Processing and Control Q1 IF 5.7
UltraLightCPS is an ultra-lightweight deep learning model designed for colon polyp segmentation in colonoscopy images. It uses an efficient encoder based on depthwise separable convolutions and an adaptive compressed multi-scale semantic fusion mechanism to capture polyp features at different scales. Its decoder combines high-level semantic information with low-level spatial details to produce accurate segmentation masks, including irregular polyp boundaries.
Authors: Shadman Sobhan, Abduz Zami, Mohiuddin Ahmed, Tanvir Mahtab Zihan, Tanvir Ahmed Khan, Aranya Saha
Conference: NCIM 2025 (Dhaka) | Publisher: IEEE Xplore
This paper presents a tuberculosis detection framework that combines lung segmentation, classification of segmented lungs, and the application of explainable AI. Initially, UNet++ is applied to segment the lung regions in chest X-rays before applying EfficientNetB0 for tuberculosis classification. To include interpretability of its predictions, LIME, SHAP, and Grad-CAM were applied.
Authors: Shadman Sobhan, Aranya Saha, Tanvir Ahmed Khan, Abduz Zami
Conference: NCIM 2025 (Dhaka) | Publisher: IEEE Xplore
Skin cancer classification from dermoscopic images is complicated by severe class imbalance; such as, benign lesions vastly outnumber malignant ones in clinical datasets. This leads to biased models that underperform on the minority classes that matter most. This paper leverages transfer learning from pre-trained CNNs combined with advanced data augmentation strategies to effectively counter this imbalance.
Authors: Abduz Zami, Shadman Sobhan, Mohiuddin Ahmed, Rakibul Islam
Conference: ICTIS 2025 (Bangkok, Thailand) | Publisher: Springer Nature
This paper proposes TIG-UNet++, a modified variant of the UNet++-L4 architecture that incorporates a thresholded version of the input image as a late-stage guide, injecting structural prior information directly into the decoding pathway. By introducing the thresholded input at a strategically chosen layer rather than at the network entry, the model preserves learned deep features while benefiting from the coarse boundary information provided by the threshold.
Authors: Abduz Zami, Shadman Sobhan, Tanvir Mahtab Zihan, Mahmudul Ahsan, Farhana Rahman, Mohiuddin Ahmed
Conference: NCIM 2025 (Dhaka) | Publisher: IEEE Xplore
Facial emotion recognition is challenged by variations in lighting, pose, and occlusion that degrade purely appearance-based models. This paper introduces a multimodal deep learning architecture that fuses landmark-based geometric features — encoding the structural relationships between facial keypoints — with spatial pixel-level features extracted by a CNN backbone.
Research works available as preprints or currently under peer review at various venues.
Authors: Shadman Sobhan, Dr. Mohammad Ariful Haque (Undergraduate Thesis)
Supervisor: Dr. Mohammad Ariful Haque | Status: Preprint (arXiv)
Standard Retrieval-Augmented Generation (RAG) systems excel at plain-text retrieval but struggle with technical documents that contain structured data such as tables and figures. My undergraduate thesis proposes a RAG pipeline that explicitly handles these elements, extracting and indexing tables and images alongside text for enhanced retrieval.