A Machine Learning Based Approach to Skin Lesion Segmentation Using Superpixels

Overview

Objectives/Goals Lesion segmentation plays an important role in the early identification and treatment of melanoma, a deadly skin cancer. Our objective was to develop an end-to-end methodology for segmenting skin lesions that was independent of dermoscopic image resolution and size in order to allow for increased efficiency and greater flexibility in implementation. Methods/Materials We employed the SLIC pre-processing algorithm to split images into smaller, spatially coherent areas called superpixels, and we extracted two feature sets that both included texture properties and distance metrics, with one set also containing RGB values and the other a color histogram. Random forest classifiers, dense neural networks, and two distinct convolutional neural networks were designed and tested to find their optimal configurations, after which we employed a post-processing, hole-filler algorithm to improve segmentation predictions. Results Our color histogram-based random forest achieved the highest accuracy of 89.83%, and our color histogram-based dense neural network obtained an accuracy of 89.01% but represents a more efficient methodology. Both of these superpixel-based segmentation techniques demonstrated comparable accuracy levels to prior studies done on a pixel level while processing 145 to 4,070 fewer pixels, making the process significantly more efficient and allowing it to operate on a wide array of images. Conclusions/Discussion We propose a pipeline that includes superpixels and our color histogram-based feature set, a random forest for machines with sufficient computing power or our dense neural network for those without, and the implementation of our post-processing algorithm to achieve efficient and high-accuracy segmentation of dermoscopic images regardless of resolution. Our results indicate that segmenting dermoscopic images using a superpixel-based approach can perform comparably to machine learning techniques on a pixel basis, even at the faster speed and that superpixels can potentially be used in a wide variety of medical image analyses to increase efficiency and flexibility while maintaining accuracy.

Competition history

• CSEF 2017 Mathematical Sciences · Entry S1502

Resources

• CSEF project summary ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Mathematics projects
• CSEF Mathematics projects
• Every CSEF 2017 project