Masters Applied Projects

Project 9

Segmentation of Colorectal Cancer with IBM Watson Machine Learning

Student: Matthew Grudza

Mentors: Dr. John Chang – Banner Health
Dr. Vikram Kodibagkar – SBHSE
Dr. Scott Beeman – SBHSE

YouTube Link: View the video link below before joining the zoom meeting

Zoom link: https://asu.zoom.us/j/82403158878

Time: 10am – 12:30pm

Abstract

Colorectal cancer is the second highest cause of cancer related deaths in the United States, responsible for about 53,200 deaths per year. Diagnosis of colorectal cancer requires time-consuming and tedious work by radiologists, analyzing computed tomography (CT) scans and manually segmenting irregular masses throughout the scan. Early detection of the cancer greatly enhances the chance of survival but proves to be difficult. A “second-observer” that can validate the radiologists’ findings could prove to be extremely beneficial for successfully identifying early-stage colorectal cancer. As knowledge in artificial intelligence (AI) advances – specifically with neural networks – AI could prove to be this second-observer to increase the accuracy of CT scans for cancer. IBM has developed an algorithm with their IBM Watson that utilizes a UNET neural network architecture to train machine learning models to segment irregular masses throughout CT scans. The algorithm can train based on different skip values that determine how many slices of the CT scan to skip while training. After analyzing the preliminary results, it was proven that by skipping more slices while training, the algorithm can sustain equivalent results while reducing the computation time to train. There were no observed differences between the skip 0, skip 1, and skip 2 segmentations. The algorithm can also incorporate ensemble learning to reduce the number of false positives. Ensemble 3 learning resulted in an 83.2% decrease in false positives from ensemble 1 learning, along with a 50% increase in false negatives. Although future improvements are necessary for clinical applications, the results are promising to show the impact machine learning can have for improving diagnosis.