Research

PhD Thesis: Advances in Deep Learning for Medical Image Segmentation and Classification

Overview

My doctoral research addressed critical challenges in medical image analysis through the development of innovative deep learning architectures and methodologies. The work focused on improving accuracy, computational efficiency, and generalization capabilities of AI models for medical diagnostics, with particular emphasis on breast ultrasound imaging and tumor detection.

Key Research Contributions

EfficientU-Net: Parameter-Optimized Medical Image Segmentation

Developed EfficientU-Net, a novel deep learning model integrating EfficientNet-B7 and atrous convolution into the U-Net architecture, optimizing breast tumor segmentation and classification in ultrasound images with reduced computational complexity and enhanced accuracy.

  • 13x reduction in parameters (1.31M vs. 17.27M in U-Net)
  • Superior segmentation of malignant tumors with irregular shapes
  • Enhanced boundary localization with adaptive receptive fields
  • 97.905% accuracy in tumor classification (benign, malignant, normal)
  • Validated on two public datasets using 5-fold cross-validation

2. UMA-Net with Adaptive Loss Functions

Developed UMA-Net, an advanced U-Net variant integrating residual connections, attention mechanisms, and atrous convolutions, enhanced by a dynamic ensemble loss function to optimize medical image segmentation across diverse datasets.

  • Residual connections and attention blocks enhance feature integration and focus on critical regions
  • Atrous convolutions enable multi-scale feature capture without compromising resolution
  • Dynamic ensemble loss (BCE, Dice, Hausdorff, Tversky) adapts weights for balanced optimization
  • Achieves superior generalization across five breast ultrasound datasets (BUET, BUSI, Mendeley, OMI, UDIAT)

3. Deep Learning and Genetic Algorithm-based Ensemble Model for Feature Selection and Classification

Developed a unified approach integrating MobileNet for feature extraction, Genetic Algorithms (GA) for feature selection, and an ensemble model for classification, optimizing medical image analysis with enhanced accuracy and efficiency.

  • MobileNet as an optimal feature extractor for medical images with minimal parameters
  • GA-based feature selection to navigate complex feature spaces effectively
  • Novel ensemble model with soft voting for robust classification decisions
  • Addresses overfitting in limited data scenarios, enhancing model generalizability
Read Full Thesis Summary

Current Research Projects

Domain-Aware Optimization for Medical Image Segmentation

Addressing the unique challenges of medical image segmentation through domain-aware optimization approaches that integrate anatomical priors, imaging physics, and clinical workflow requirements into automated machine learning pipelines.

Research Focus:

  • Develop optimization frameworks that incorporate domain-specific constraints and anatomical priors
  • Create efficient strategies for limited medical imaging data scenarios
  • Design multi-objective optimization balancing accuracy, uncertainty quantification, and computational efficiency
  • Ensure compliance with clinical requirements and regulatory standards
  • Validate across diverse medical imaging modalities and clinical applications

Saliency-Guided AttentionNet for Breast Ultrasound Classification

A novel dual-branch deep learning architecture that addresses the challenges of breast ultrasound classification by explicitly modeling both lesion-specific and peritumoral tissue features through saliency-guided attention mechanisms.

Key Innovations:

  • Dual-branch architecture with shared MobileNet backbone for efficient feature extraction
  • Grad-CAM based saliency guidance for explicit foreground-background decomposition
  • Adaptive channel-wise attention for dynamic feature fusion
  • Validated on 1,551 images across 5 public datasets with 90.51% accuracy
  • Robust cross-dataset validation with 78.46% accuracy on held-out sets

Fuzzy Similarity-Driven Loss for Medical Image Segmentation

A novel Fuzzy Rough Set (FRS) loss function that enhances boundary delineation in medical images by effectively handling uncertainty and ambiguity in lesion boundaries through fuzzy similarity measures and rough set approximations.

Key Contributions:

  • Novel FRS loss function based on fuzzy rough set theory for precise boundary delineation
  • Enhanced sensitivity to uncertain predictions through new fuzzy similarity function
  • Effective handling of boundary ambiguities using lower and upper approximations
  • Validated across multiple modalities (ultrasound, MRI, CT) and clinical applications
  • Reduced computational complexity compared to conventional loss functions

Uncertainty-Aware Explainable Deep Learning for Medical Image Segmentation

A comprehensive framework integrating pixel-level uncertainty estimation with multi-modal explainability techniques to enhance clinical trustworthiness and safety in AI-assisted medical image analysis.

Key Features:

  • Monte Carlo dropout and Bayesian layers for robust uncertainty estimation
  • Multi-modal explainability (Grad-CAM, attention mechanisms, saliency maps)
  • Mean Dice coefficient of 0.89±0.03 across multiple datasets
  • 92% segmentation failure detection with only 15% manual review rate