Digital Image Processing Class Final Project Spring 2009: Intelligent Data Capture of X-Ray Images for Use on Mobile Phones

This semester, I took the graduate course 6.344 Digital Image Processing. Our final project was to apply the image processing techniques we learned that semester to an interesting problem. It was pretty open-ended, so I decided to do a project that would also be useful for Moca: intelligent data capture on phones of x-ray images.


Background: One application for Moca is for the healthcare workers to capture images of the patient’s chest x-rays to detect tuberculosis, so that they can be sent to the radiologist in the city hospital for review. Otherwise, the paper x-ray would have to be physically trucked over to the hospital, which only occurs after enough x-rays pile enough to make the expensive trip worth it. The key, though, is for the healthcare worker to take a high enough quality image for the radiologist to be able to diagnose the patient properly for tuberculosis or not.


However, because healthcare workers aren’t trained to be experts, sometimes they naively take a picture that is not too small, blurred, under or over exposed, or they might cut off a portion of the image such as the written information on the x-ray. (This is reported by radiologists who have worked in the United States and in the Philippines). This makes the x-ray unreadable but by the time the problem in the image is seen by the radiologist, the patient has long left the rural health clinic and it’s not possible to take the image again.

Hence, we need a way to intelligently assess the quality of the image right when it is captured on the phone by the healthcare worker. We want to be able to tell them immediately if the image is (1) cut off or rotated, (2) under or over exposed, or (3) blurred. If it is, the application running on the phone would recommend them to immediately retake the image. In order to test these algorithms properly, I will use a set of x-ray images taken by a Google Android cell phone and provided to me by the University of Philippines National Telehealth Center.

Image Blurred


Image Cut Off


Image under exposed


As a result of this project, I was able to come up with a Matlab GUI that took as input an x-ray image and then output a result to a healthcare worker that determined whether the image of the x-ray was good enough quality to go onto a radiologist or if the healthcare worker should take the image again because of 1 of 3 reasons. When we actually integrate this with Moca, it will basically sit on the Android phone, wait for the healthcare worker to take an image of the x-ray with the phone, do some image processing, and then return a recommendation to the healthcare worker right then and there.


I used three different types of algorithms: blur detection, under/over exposure detection, and corner detection (to detect if the image was cut off on the edges). There is also the case where the healthcare worker doesn't take the advice of the automated image analyzer and doesn't retake the photo. Then we want to do what we can to provide the radiologist with the best image quality possible. I tried to cover this case as best as I could by cropping the image (excess x-ray film borders), sharpening the image, and performing histogram equalization (to even out the bright/dark regions) on the image, and these modifications to the image are seen in the above GUI.