Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Classification of colorectal polyps detected during standard colonoscopy as adenomatous or hyperplastic using image analysis and machine learning algorithms

Patikos Ioannis

Full record

---

Summary

Nowadays, colorectal cancer (CRC) is one of the most frequent causes of cancer fa-tality worldwide. Researches have shown CRC’s intimate relation to colorectal polyps. The vast majority of all gastrointestinal carcinomas are considered to origi-nate from adenomatous polyps and as a result, their early detection could prevent their transformation to cancer. Hence, new methods that are trying to enhance the adenoma detection rate (ADR) are being researched and developed. The employ-ment of artificial intelligence (AI) techniques, like deep learning, and especially con-volutional neural networks (CNNs), helps to identify cancerous tumors and colonic polyps. The CNN architecture is well-suited by design to provide beneficial solutions, including polyp detection and classification. On that account, four different CNN models have been implemented in the current thesis. The first three are dealing with the polyp detection problem, while the fourth one performs polyp’s classification as ‘‘adenomatous’’ and ‘’hyperplastic’’. In both tasks, a binary classification takes place. In the first case, image data are classified into "polyp" and "healthy" categories and in the second case into "adenomas" and "hyperplastic". A combination of various im-proving techniques has been applied in the first three models to see how they affect the performance of the CNNs. These methods consist of: Image Data Augmentation, the Dropout Regularization technique and the Transfer Learning technique. In the CNN model for polyp classification as adenomatous and hyperplastic polyps, the data were related to the values of the GLCM texture features of the images. The data, used in this thesis, were collected retrospectively from the extensive personal ar-chive of doctor Konstantinos Patikos. From a total of 750 patients, 1576 images were collected; 798 contain polyps and 778 depict a healthy colon. The 798 images with polyps are separated into two categories; 424 pictures with adenomatous polyps and 374 pictures with hyperplastic polyps. All the data come from standard colonoscope, which uses white light for the inspection of the bowel wall. The images are not very uniform, as they tend to differ in zoom, focus or coloration. The training and testing datasets for the polyp detection task contain 1470 and 106 samples respectively. The training and testing datasets for the polyp classification problem consist of 170 and 34 samples each. Performance metrics like Accuracy, Sensitivity, and Specificity were measured to evaluate the effectiveness of the proposed models. The most effi-cient scenario that dealt with the polyp detection task scored 92.2% accuracy, 94.4% sensitivity, 90.6% specificity, 90.9% precision, 9.4% FPR and 5.6% FNR over unseen data. The only scenario that confronted the polyp classification problem achieved 85% accuracy, 88.8% sensitivity, 81.3% specificity, 84.2% precision, 18.7% FPR and 11.1% FNR.

Available Files

Services