The project is an implementation of a computerized system for early analysis of beauty spots in order to detect suspicious spots such systems which exist today
Abstract
The project is an implementation of a computerized system for eary analysis of beuty spots in order to detect suspisous spots such systems which exist today, tries to help the expert doctors in their final analysis and decision making our system tries to help earlier stage of the analysis proccess which is usually done by a family doctor.
The system perofrms the following stages:
- Segmantation of a given image of a beauty-spot
- Analysis and Classification of the spot according to ABCDE parameters + additional paramater developed by us.
- Updating a “Learning Data-base” for further spots analysis
Background & Motivation
Melanoma is one of the most common and dangerous types of cancer known today. Early detection of melanoma can increase the chances of recovery and even prevent the spread of the cancer. Nowadays, the primary analysis is done by a simple examination by a family doctor. In this project, we check the possibility of replacing the primary examination, which is being done by a doctor, with a computerized system. This method will save a large amount of time by using image processing technologies. The system includes, in addition to image processing, a learning algorithm which uses a data-base that can be updated while the system is in use. Using this concept, the systems success rate will increase along with the usage of the system. Our system bases on the known ABCDE parameters of beauty-spot analysis along with additional parameter from the field of image processing.
The solution
In general, the system’s analysis flow can be described in the following flow-chart:
The classification is done using a learning system (implemented using SVM algorithm)
The Parameters which being analyzed are the following:
Results
The two major statistics which gives a feeling about the system performance are:
1. Correct Rate: How many correct classifications were done comparing to the total number of classifications
2. False Negative: How many dangerous spots were identified as not-dangerous, comparing to the total number of classifications.
More statistic can be found in the project book at the links below
1. Correct Rate
2. False Negative
Conclusions
- Success rate grows along with the size of the training group (as expected)
- From a training group size of ~50% of the DB, the success rate and the false-negative rate are very good
- Sticker normalization has minor effect on the results (we believe that it is due to the very small DB)
- Larger DB will enable higher performance and will assist in finding new parameters which human eye can’t detect