Skip to content

Latest commit

 

History

History
79 lines (55 loc) · 5.36 KB

File metadata and controls

79 lines (55 loc) · 5.36 KB

Dermoscopic Image Analysis for Lesion Detection using Few-Shot Learning

Model building, experiments, references, and source code for research w theork on skin image analysis that draws on meta-learning to improve performance in low data and imbalanced data regimes.

This research was carried out in collaboration with the Machine Learning Research Group (MLRG) and the Departments of Computer Science and Information Technology at SSN College of Engineering, Anna University, India.

Quick Links

The following references will aid in reproducing this implementation, and to extend the experiments presented in the paper for further analyses.

Cite Us

Link to the Research Paper (preprint version).

If you find our work useful in your research, please cite us:

@article{https://doi.org/10.48550/arxiv.2210.16954,
  doi = {10.48550/ARXIV.2210.16954},  
  url = {https://arxiv.org/abs/2210.16954},  
  author = {Desingu, Karthik and P., Mirunalini and Chandrabose, Aravindan},  
  keywords = {Computer Vision and Pattern Recognition (cs.CV), Artificial Intelligence (cs.AI), Machine Learning (cs.LG), FOS: Computer and information sciences, FOS: Computer and information sciences},  
  title = {Few-Shot Classification of Skin Lesions from Dermoscopic Images by Meta-Learning Representative Embeddings},  
  publisher = {arXiv},  
  year = {2022},  
  copyright = {Creative Commons Attribution 4.0 International}
}

Open-source datasets used for evaluation.

Motivation

  • Annotated images and ground truth for the diagnosis of rare and novel diseases are scarce. This is expected to prevail, considering the small number of affected patient population and limited specialized clinical expertise to annotate images.
  • Further, the frequently occurring long-tailed class dataset distributions in skin lesion and other disease classification datasets cause conventional training approaches to lead to poor generalization due to biased class priors.
  • Few-shot learning, and meta-learning in general, aim to overcome these issues by attempting to perform well in low data regimes.

Proposed Embedding Network & Base-Learner Approach for Meta-Learning

This work focuses on improving meta-learning for the characterization of lesion types from dermoscopic images.
Specifically, it proposes a two-stage training and inference approach,

  • A baseline supervised learner on the meta-training set that allows a network to learn highly representative and generalizable feature embeddings for images, that are readily transferable to new few-shot learning tasks.

  • Positing that a representative feature embedding can be more effective than complex meta-learning algorithms, a simple classifier is trained atop these representations for downstream classification into lesion types.

Key References

  • Proposes the use of meta-learning techniques for efficient model adaptation for extremely low-data scenarios
  • Applies Group equivariant convolutions (G-convolutions) in place of the normal spatial convolution filters
  • Two network implementations:
    • Reptile: Gradient-based meta-learning
    • Prototypical networks using Euclidean Distance
  • Evaluated on ISIC 2018, Derm7pt and SD-198 datasets
  • Outperforms DAML on ISIC 2018
  • Implementation Code NOT available
  • The paper that proposed Relation Networks for Few-Shot Learning.
  • The paper that proposed Protoypical Networks for Few-Shot Learning.
  • An improvement of Prototypical Networks, by generating query-specific prototypes and thus results in local and distribution-aware metric
  • Sets different temperature for different categories to penalize query samples that are not close enough to their belonging categories.