Research Background

Research Focus

My research focuses on improving few-shot learning for medical image classification by enhancing Prototypical Networks within a meta-learning framework. The work addresses the critical challenge of data scarcity in clinical AI, where medical datasets often contain very limited labeled samples. 

Specifically, the research aims to design more robust and discriminative prototype-based learning mechanisms by improving how prototypes are constructed, how features are represented and refined, and how episodic training tasks are sampled. 

The key contributions of the research include:

• Prototype Construction Improvement: Developing influence-based reweighting strategies so that more informative support samples contribute more strongly to class prototypes.

• Feature Refinement: Selecting and emphasizing the most discriminative feature maps and channels to improve prototype quality.

• Domain-Specific Feature Integration: Incorporating texture-aware representations (e.g., wavelet-based features) to capture subtle morphological patterns in medical images.

• Optimized Episodic Sampling: Designing adaptive and metadata-guided sampling strategies to create more informative few-shot learning tasks.

Overall, the research seeks to build efficient, robust and interpretable few-shot learning frameworks that can accurately classify medical images even when only a few labeled examples are available, thereby improving the applicability of AI in real-world healthcare settings

Domains Addressed

In my phd work I have tried to solve few shot classification problems in various medical imaging domain which are displayed below :

• Dermatoscopic Images

1.  Derm7pt Dataset

2.  SD-198 Dataset

3.  ISIC-2018 Dataset

4.  ISIC-2019 Dataset

• Breast Cancer Histopathological Image Dataset(BreakHis Dataset)

• Blood Cell MNIST Dataset

• Pathology MNIST Dataset

• Chest XRay MNIST Dataset

• DermaMNIST Dataset

• OrganAMNIST Dataset

• NIH Chest XRay Dataset

• Diabetic Retinopathy Dataset