Weakly Supervised Segmentation of Breast Cancer Ultrasound Images
Click the Poster to View Full Screen, Right click to save image
A'nya Carr
College:
The Dorothy and George Hennings College of Science, Mathematics, and Technology
Major:
Computer Science
Faculty Research Advisor(s):
Kuan Huang
Abstract:
The purpose of this project was to generate labels for breast cancer tumors in ultrasounds by exploring deep-learning methods of Weakly Supervised Segmentation. Syed, S., Anderssen, K.E., Stormo, S.K. et al. in their study "Weakly supervised semantic segmentation for MRI: exploring the advantages and disadvantages of class activation maps for biological image segmentation with soft boundaries" describe Weakly Supervised Segmentation as a method "...where image-level labels and class activation maps (CAM) can detect discriminative regions for specific class objects." This project's experiment was conducted by using a dataset of ultrasound images containing benign and malignant tumors and their Class Activation Maps using a traditional ResNet-50 model, and a custom CAM method introduced by Jie Qin, Jie Wu, Xuefeng Xiao, Lujun Li, and Xingang Wang. An innovative CAM method called AMR was introduced in "Activation Modulation and Recalibration Scheme for Weakly Supervised Semantic Segmentation" (AMR) by Jie Qin, Jie Wu, Xuefeng Xiao, Lujun Li, and Xingang Wang, to isolate the surrounding regions of an object in an image and generate a mask of its location. This saves researchers and doctors time meticulously annotating each image in a dataset.
In training, traditional ResNet-50 and AMR were trained for 20 epochs, a batch size of 12, and a learning rate of .001. Then, traditional ResNet-50 CAMs were generated and compared to the AMR CAMs generated. The product was evaluated by the accuracy of pinpointing the tumor’s location using Intersection over Union (IOU). IOU serves as a metric to gauge the accuracy of the CAM in object detection. With an ultrasound image sample containing a benign tumor, the traditional CAM scored a .03, signifying that the model failed at locating the tumor. Using the same image, the AMR CAM scored .3, signifying that the model vaguely knows the location of the tumor. A good score would be around .7 and .9, meaning more research and analysis must be done to improve the model’s detection of benign tumors.
Different results were noticed in a sample containing a malignant tumor, with a traditional ResNet-50 CAM scoring of .08 and AMR CAM scoring of .63. The traditional CAM model failed at detecting the malignant tumor. With AMR CAM, the model nearly has a good accuracy of detecting the malignant tumor. Further research is needed to improve these object detection scores and to explore the effectiveness of Weakly Supervised Segmentation on other datasets of ultrasounds. Nonetheless, the results show promise in improving the process of detecting breast tumors in ultrasounds for the medical industry.