Intelligent Information Detection in Medical Images

Abstract

This Project report describes our objective of INTELLIGENT INFORMATION DETECTION IN MEDICAL IMAGES, where I tried to diagnose lung cancer using Digital CT scans by building a classifier. I have implemented supervised classifiers for cancer detection. Seven supervised learning classifiers implemented were K-Nearest Neighbors, Support Vector Machine, Decision Tree Multinomial Naïve Bayes, Stochastic Gradient Descent, Random Forrest, and Multi-Layer Perceptron. Images obtained from digital CT scans aren’t always noise free, which could greatly hamper the detection process. In order to solve the problem of noise in images and to extract features from the given images, I have implemented various image processing techniques. These techniques varied from using Gaussian mean to remove noise from the images to using Otsu’s thresholding method for converting the gray-scale images to binary images. Even though thresholding is quite a good method for conversion of gray-scale image into binary image, it still leaves behind gaps in the image which could mislead the classifier to wrongly classify an image. While these models were successfully implemented and evaluated on a small, segregated, consistent dataset; inconsistencies due to use of machine from different manufacturers, used on patients from different positions and angles in our dataset and in the real world application presented a real challenge in achieving our objective completely.