04/16/2021
By Qilei Chen

Title: Deep Learning for Lesion Detection on Medical Retinal Images

Ph.D. Candidate: Qilei Chen
Time: Friday, April 30, 2021, 10 a.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
  • Hengyong Yu (member), Professor, Department of Electrical and Computer Engineering

Abstract:
The past decade has witnessed the rise of Convolutional Neural Networks (CNNs) in deep learning. The CNN models have succeeded in a variety of general application domains such as Computer Vision (CV), Natural Language Processing (NLP), Automatic Speech Recognition (ASR), etc. CNNs' great achievements illuminate their potentials in the medical and health care fields. One of the major applications is the design and evaluation of CNN-based lesion detection techniques for early diagnosis and continuous monitoring of patients suffering from eye diseases. In this dissertation, we conduct an in-depth study on CNN-based lesion detection in medical retinal images for diabetic retinopathy (DR) and retinopathy of prematurity (ROP).

Lesions in the human retina are damages or abnormal tissues. Many of them are early manifestations of fatal diseases, which can lead to blindness in certain cases. Thus, lesion screening is a crucial process at early stages, which can significantly increase the cure rate. Retinal imaging is a traditional technology that helps optometrist to assess the health of the human eye using a high-resolution camera to take a picture of the back of the eye. For patients with DR or ROP, retinal images can be compared side-by-side over time to monitor the eye health and detect subtle lesions. Visual lesion detection on retinal images has experienced significant improvements in computer-aided system with deep learning in recent years. However, due to the variety of lesion types and complex normal anatomical structures, automatic detection of lesions on retinal images remains a challenging task.

First, retinal lesions at early stages are usually very sparse and small. Thus, the visual size of lesion instances on the high-resolution retinal images are usually tiny, making them difficult to be detected. The second challenge is the incompleteness of manual annotations in the medical retinal image dataset. This is because it is laborious process for medical experts to fully label a large number of medical images in the dataset. Some of the small lesions may be missed in the process. As a result, only a subset of all lesion instances is marked and the lesions without manual annotations on the images will be mistakenly counted as negatives when training a CNN model. The third challenge is that that visual lesion instances provide only partial information that is not enough for final medical diagnose, such as the stage grading in ROP.

To address the first challenge, building upon the architecture of state-of-the-art CNN detection models, we develop two novel detection models for effective small object detection. One of them adopts the large-size feature pyramid network (LFPN) that preserves the details of local regions on the retinal images. The other model enhances the result with deeper feature pyramid network, which is more device-friendly. For the second challenge, we introduce an iterative training algorithm based on pseudo-labeling, a semi-supervised method through which a considerable number of unlabeled lesion instances can be discovered to boost the performance of the lesion detector. Furthermore, to take advantage of the local lesion detection results and make up for the incompleteness of local information in ROP stage grading, we propose a local-global information combination mechanism for ROP stage classification, motivated by the fact that the formation and shape of a demarcation line in the retina is the distinguishing feature between earlier ROP stages. Finally, to make the process of early screening more effective, we propose a multi-model system to analysis the DR images, through which the quality, the stage and more useful information on the image can be revealed for further diagnosis.

Our current and future work will investigate the model optimization problem about how to make the prediction more accurate and use the region information for localization on the challenging DB and ROP image datasets.