11/24/2021
By Dechun Wang

The Kennedy College of Sciences, Department of Computer Science, invites you to doctoral dissertation defense by Dechun Wang on "Deep Learning for Colorectal Polyp Detection and Instance Segmentation on Colonoscopy."

Ph.D. Candidate: Dechun Wang
Time: Monday, Dec. 6, 2021, 2 p.m.
Location: This will be a virtual defense via Zoom.

Committee Members:

  • Yu Cao (advisor), Professor, Department of Computer Science
  • Benyuan Liu (advisor), Professor, Department of Computer Science
  • Bo Hatfield (member), Chair, Department of Computer Science, Salem State University


Abstract:

With recent advancements in machine learning, Convolutional Neural Networks (CNNs) have shown their potential on image processing. Among their achievements are a variety of intellectual tasks in a wide range of general areas, such as Computer Vision and Natural Language Processing. With substantial

work on developing the base model on natural image domains, the achievements of CNNS also convey their potentials in the medical domain. A number of tasks, such as classification, object detection, and segmentation, have gained prominence in recent years due to their widespread applications. In this dissertation, we will investigate and develop CNN-based object detector and instance segmentation techniques for colorectal polyp detection.

In the colon or rectum, polyps are abnormal growths that are at high risk of developing into colorectal cancer (CRC), one of the most serious health issues worldwide and the third most common type of cancer with high mortality. It is reported that 52,980 people will die from CRC in 2021. Colonoscopy has been shown to be a reliable method for detecting and removing polyps in the colon early on, thereby preventing colorectal cancer. However, according to Eufkens et al.’s study, 22%-28% of polyps and 20%-24% of adenomas are missed in colonoscopy due to various human factors. Therefore, there is a critical need for an efficient and accurate computer-aided colorectal polyp detection system that can assist physicians with localizing the polyps during a colonoscopy. However, due to the variety of polyps in terms of shape, size, color, and texture, automatic detection of colorectal polyps is a challenging task. To help physicians detect polyps in colonoscopy procedures, we devised a special anchor-free based colorectal polyps detection for the general polyp detection purpose. This anchor-free based polyps detector can detect and localize polyps without using predefined anchors. To improve the model's ability to detect small polyps, we devised a Context Enhancement Module to utilize the context information at the feature level. Furthermore, by incorporating contextual information, false positives caused by artifacts, such as fluid, debris, bubbles, reflections, specularity, contrast, saturation, and medical instruments are reduced. In addition, we utilized Cosine Ground truth Projection to improve the overall recall rate. As a result, we achieved the best performance on the mainstream public dataset (CVC-Clinic dataset, ETIS-LARIB dataset) with a real-time speed. The aforementioned approaches were mainly designed to detect objects within bounding boxes, however, the need to find polyp boundaries is constantly increasing. To address this, we devised a new auto-encoder based method with a unified framework that generates instance-level mask in a single-stage fashion. Our framework encodes mask into a compact representation vector, which allows us to incorporate the instance segmentation with one-stage bounding-box detectors in a simple yet effective way.

We also investigate the performance degradation in colonoscopy caused by the significant difference between the image of the real-world colonoscopy procedure and the public dataset. To this end, we collected a private dataset that is representative of realistic colonoscopy operations, trained and evaluated our models with the new datasetset.

Our current and future work will continue to investigate the model optimization method and improve the model's overall performance and robustness in a dataset that represents real-world operations.