Skin Lesion Characterization

Analysis of Dermoscopic Skin Lesion Images to segment lesion regions and characterize the lesion type using deep adversarial learning (conditional GANs) and EfficientNet-based classifiers.

Quick links

• Online research article [Springer Nature, HTML].
• Published manuscript [PDF].
• Data directory organization [Markdown] (internal reference).

Cite Us

Link to the Research Paper.

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

@article{kd2024conditional,
  title={Conditional adversarial segmentation and deep learning approach for skin lesion sub-typing from dermoscopic images},
  author={Mirunalini, P and Desingu, Karthik and Aswatha, S and Deepika, R and Deepika, V and Jaisakthi, SM},
  journal={Neural Computing and Applications},
  pages={1--19},
  year={2024},
  publisher={Springer}
}

Study design overview

The different choices of data balancing and segmentation approaches used to augment the skin lesion characterization pipeline are presented in the following schematic.

Dataset balancing and lesion segmentation

The dataset balancing analysis and lesion segmentation network is contained in notebooks. The python notebooks are sequentially numbered; simply run them serially to reproduce the presented experiments.

Scaling experiments for classification

The classification architectures for classifiers are scripted in src/classifiers. The preprocessing workflow used to prepare the dataset is contained in model-building.

• Load the appropriate classifier driver function, say experiment_effnetb6 (refer to the .py files and import statements), in src/run.py by importing them.
• Set up the data path.
• Call the driver function in src/run.py, and execute the script with python run.py.

---

Environment setup

Set up the execution environment using the requirement files.

• Requirements for setting up conda environment are contained in dep-file-conda.yml.
• Requirements for setting up using pip installations (not recommended) are contained in dep-file-pip.txt.

Note

The proprietary segmentation dataset generated with the help of expert pathologists used will be released upon reasonable request after the research manuscript is published.