Future Work

While we are satisfied with our final project results, we would suggest the following improvements to be made outside the time constraints of a semester-long project to further improve our algorithm's accuracy.

Image Courtesy of: Kabir Deol and https://www.isic-archive.com/#!/topWithHeader/onlyHeaderTop/gallery?filter=%5B%5D

Further Improve Segmentation

Due to the limitations of Otsu's method mentioned in Segmentation, our algorithm has difficulties segmenting a melanoma region when an image has poor lighting, the skin is of darker pigment, and when the image contains the edge of a body part (as shown to left). To improve our algorithms accuracy for these situations, we would implement a variable threshold that adjusts based on a certain image rather than relying on the threshold determined by Otsu's method.

Filtering Hair

The accuracy of our algorithm decreased when dark hair was covering the lesion area we were analyzing. To address this problem, we would develop a DullRazor filter in Matlab that would classify hair pixels in an image by identifying their characteristic shape. Next, the filter would use linear interpolation on the hair pixels and a Gaussian filter to output a new image with hair removed.

Image Courtesy of: http://www.dermweb.com/dull_razor/

Non-Dermoscopic Images

We designed our algorithm to work on dermoscopic images of skin, which are images taken that contain a singular lesion area of interest. To improve the functionality of our algorithm, we would adapt our code to work on images containing a large region of skin with multiple potential melanomas. The algorithm would identify the location of potential melanomas and classify an area of interest surrounding the location. Finally, the algorithm will run our current ABC classification method on each area of interest and output the number of identified melanoma.

Image Courtesy of: https://www.researchgate.net/figure/Distant-top-and-closer-lower-right-and-lower-left-views-of-the-clinical-presentation_fig2_319892458