karthik-d / lesion-characterization-using-cgan

Analysis of Skin Lesion Images to segment lesion regions and classify lesion type using adversarial deep learning.

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

• Manuscript draft [PDF].
• Data directory organization [Markdown] (internal reference).

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.

Dataset balancing and lesion segmentation

The dataset balancing analysis and lesion segmentation network is contained in notebooks as sequentially numbered python notebooks.

Scaling experiments for classification

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

• Load the appropriate classifier drive function, say experiment_effnetb6, 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.

Note

The proprietary dataset used will be released after the research manuscript is published upon request.