Home > Events > Oral Candidacy - Yizhe Zhang

Oral Candidacy - Yizhe Zhang

Start: 8/11/2017 at 2:00PM
End: 8/11/2017 at 5:00PM
Location: 100 Stinson Remick
Attendees: Faculty and students are welcome to attend the presentation portion of the defense.
Add to calendar:
iCal vCal

Yizhe Zhang

Oral Candidacy

August 11, 2017              2:00 pm           100 Stinson Remcik

Adviser:  Dr. Danny Chen

Committee Members:

Dr. Sharon Hu          Dr. Walter Scheirer          Dr. Yiyu Shi

Title: 

NEW ALGORITHMS AND DEEP LEARNING MODELS FOR OBJECT DETECTION AND SEGMENTATION IN BIOMEDICAL IMAGES

Abstract: 

Object detection and segmentation are fundamental problems in computer vision research field. In digital images, a detection task is to detect objects of interests in an output form of center coordinates and/or bounding boxes. A segmentation task is to segment objects of interests to obtain their outer contours/boundaries. Both detection and segmentation are essential steps for biomedical image analysis and computer-aided diagnosis since they can provide information and insights for quantitative biological researches and disease studies.

Before the era of deep learning, detection and segmentation problems are usually tackled by using handcrafted features. In the last five years, deep learning models have achieved many successes in biomedical image analysis. Our preliminary work has been focused on both algorithmic oriented methods (non-deep learning methods) and deep learning based methods for object detection and segmentation in biomedical images. 

For algorithmic based methods, we developed a seeding-searching-ensemble method which works on generating segmentation region proposals. This method is good at preserving weak object boundaries and providing less number of over-segmented region proposals. We demonstrated the advantage of this method in gland segmentation in H&E stained histology images.

For deep learning models, we developed a coarse-to-fine fully convolutional network. This model is designed to incrementally learn segmentation knowledge (mapping functions) from non-expert level to expert level. We demonstrated the advantage of this model in lymph node segmentation in ultrasound images. In biomedical practice, it is often the case that only limited annotated data is available for model training. Unannotated images are usually abundant. We developed a deep adversarial network model for biomedical image segmentation. The proposed adversarial network can help to attain better segmentation results for unannotated images.

Our future work is to combine algorithmic based method (e.g. superpixels generation algorithms, computational geometry algorithms) and deep learning methods to obtain better object detection and segmentation models for medical and biomedical image analysis.