07/13/2022
By Dechun Wang

The Kennedy College of Sciences, Department of Computer Science, invites you to attend a Doctoral Dissertation Proposal defense by Dechun Wang on “Deep Learning for Colorectal Polyp Detection and Instance Segmentation on Colonoscopy.”

Ph.D. Candidate: Dechun Wang
Date: Monday, July 25, 2022
Time: 11 a.m. Eastern Time
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 work, I designed an automated real-time polyp detection system based on deep CNNs. A deep CNN can be easily applied to an endoscope because endoscope produces a standard video stream.

In the colon or rectum, polyps are abnormal growths that are at high risk of developing into colorectal cancer (CRC), which is one of the most serious health issues worldwide and the third most common types of cancers with a high mortality. It was reported that 52,980 people died from CRC in 2021. Colonoscopy has been shown to be a reliable method for detecting and removing polyps in the colon if found early, 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 errors. Therefore, there is a critical need for an efficient and accurate computer-aided colorectal polyp detection system that can assist physicians with identifying the polyps during a colonoscopy. However, due to the variety of polyps in terms of shape, size, color, and texture, computer-aided detection of colorectal polyps is a challenging task.

Throughout my Ph.D. research, I developed a CNN-based computer-aided colorectal polyp detection system that consists of an anchor-free polyp detection model (AFP-Net) and an instance segmentation model (AFP-Mask) for colorectal polyp detection. The system has two advantages: (1) It has excellent performance on detecting fleeting polyps and subtle polyps. (2) It has a very low false-positive rate. Thanks to its low false-positive rate and high recall rate, the system was deployed in many endoscope centers to assist physicians with identifying polyps during colonoscopy procedures.

The AFP-Net is an anchor-free based colorectal polyp detection model for the polyp detection purpose. AFP-Net can detect and localize polyps without using predefined anchors. To improve model's ability to detect small polyps, I devised a Context Enhancement Module to utilize the context information at the feature level. This module incorporated contextual information into the AFP-Net to further reduce the false positives caused by artifacts, such as fluid, debris, bubbles, reflections, specularity, contrast, saturation, or medical instruments. In addition to these, I also devised the Cosine Ground Truth Projection to improve the overall recall rate. As a result, the system achieved the best performance on several mainstream public dataset (CVC-Clinic dataset, ETIS-LARIB dataset) in a real-time speed.

While AFP-Net model was mainly used to detect objects with bounding boxes, during the deployment of the system, to find polyp boundaries was increasingly desired. To address this necessity, I designed and implemented a new auto-encoder based model, namely AFP-Mask, to generate instance-level mask in a single-stage fashion. AFP-Mask encodes mask into a compact representation vector, which allows it to incorporate the instance segmentation with one-stage bounding-box detectors in a simple yet effective way. Compared to other instance segmentation models, AFP-Mask generates a mask that can smoothly trace the boundary of the polyp. Experimental results showed that the AFP-Mask model outperforms previous approaches on instance segmentation in terms of mIoU and DICE score while still maintaining a real-time speed.

Even though the detection system achieved outstanding performance on mainstream public dataset, it showed a lagging when applied to the data collected during real-world colonoscopy procedures. To close this gap, a large-scale polyp private dataset was collected. The dataset contains 13,104 images with object-level bounding boxes and segmentation. Compared to public dataset, this rich private dataset contains various type of polyps in terms of size, shape, and texture. This private dataset also contained data collected from three different major endoscopy manufactures, representing wider ranges of artifacts. To further challenge the system, the private dataset contained some polyps that were annotated in relatively blurry images caused by camera motion, out-of-focus, or water flushing during a colonoscopy procedure. All these factors contributed to a more challenging and realistic scenario to test the system for polyp detection performance. In comparison with the public dataset, the polyp detection system trained using this rich real-world dataset demonstrated improved performance.