TASK
Our team is working with a dataset from Mammoimage.org that includes mammogram X-rays of various types, including normal, asymmetrical, malignant, benign, and calcified images. Using MATLAB, we've processed these images to help us extract distinguishing features, and have written an algorithm that classifies malignant tumor images from normal, tumor-free ones with an average accuracy of 96.65%.
After removing some outlier images, the final dataset consists of 204 normal images and 53 malignant tumor ones.
CHALLENGES
Medical images are complicated, patient-specific, and often noisy, so it's difficult to extract quantitative features that distinguish a healthy image from one showing a tumor. Mammograms are particularly difficult because breast tissue density differs between patients, and some tumors can't be easily delineated from the background.
For the mammogram images we're working with, the average intensity is not different between normal images and tumor images -- the tumor images just have small, dense regions of high intensity. Another issue specific to our dataset is that background noise is not consistent between image, presenting a challenge for pre-processing. The tumor size and location also varies between malignant images. This rules out image processing techniques that isolate tumors based on their spatial references.
From tackling this project, we've learned that medical images present a uniquely difficult challenge in terms of image processing and machine learning. We've tried to find a combination of image processing and feature extraction techniques to identify dense, cancerous regions.