PhD Seminar Presentation by Yahya Alzahrani:"Deep Learning Approach for Breast Ultrasound Image Segmentation "

 

Date: Tuesday June 15th 2021 

Time: 11:00am-1:00pm 

Meeting URL:  https://zoom.us/j/98075338548?from=addon 

Passcode:    If interested in attending this event, contact the Graduate Secretary at csgradinfo@uwindsor.ca

 

In the past decade, more options for segmenting breast cancers have emerged as a result of recent improvements in computer-aided diagnosis (CAD) technologies. Breast cancer has been a leading cause of mortality for women globally. Offering a good diagnosis in the early stages of breast cancer using the widely available Ultrasound images is an efficient strategy to reduce breast cancer. However, it remains a challenging task due to the limitations of Ultrasound images such as intensity inhomogeneity and class imbalance. Encoder-Decoder based convolution neural networks have proven their efficiency in segmentation task in different imaging modalities. To address the above-mentioned downsides of ultrasound images, we proposed a modified U-Net architecture equipped with pre-trained inception residual encoder for breast tumor segmentation in breast ultrasound (BUS) images. we employed the inception blocks to increase the depth of the network. Our proposed solution consists of a preprocessing stage, feature extraction based on various inception blocks and a plain U-Net decoder. Two publically available datasets were utilized in our work named BUSI and UDIAT. Our proposed approach has proven to be promising, as our results show improved performance over the existing U-Net architecture, as well as the more recent DAL and SK-U-Net models. In particular, our model achieved a Dice Coefficient and IOU of 0.94 and 0.90 on UDIAT, respectively, and 0.89 and 0.80 on BUSI, respectively. 

 

Keywords:  BUS Images, Segmentation and Classification, Neural Networks 

 

External Reader: Dr. Esam Abdel-Raheem 

Internal Reader: Dr. Christie Ezeife  

Internal Reader: Dr. Dan Wu 

Advisor: Dr. Boubakeur Boufama