Using Salience and Convolutional Neural Networks to Segment Breast Tumors from Surrounding Breast Tissue Open Access
Downloadable ContentDownload PDF
Breast cancer occurs when healthy breast tissue cells are taken over by cancerous cells, and metastatic cancer is cancer that has spread from the original site of the tumor, increasing the mortality rate of the patient. One of every 8 women will develop metastatic breast cancer in her lifetime, and 39% of women with breast cancer will die from the disease. Breast cancer is the most common cancer in the United States after skin cancer. In this work, breast tumors are segmented from healthy breast tissue to create a clearer image of the breast cancer for treatment. Having a clearer image is important as surgeons need to be able to remove the tumor, leaving healthy tissue undamaged. Additionally, the software could be used to help train radiologists. The images used here have been contrast-enhanced to increase the visibility of vascularity and to correspond to truth images, which are segmentations of the tumor. This thesis will segment each MRI image by using salience to create a segmentation map, a convolutional neural network to create a segmentation map, and the two programs in combination to create the final segmentation map. The various combinations will then be compared to determine which was best able to segment the tumor region. To segment the image, the original MRI images are first passed through a salience program. Here we discuss many types of salience programs, using a comparison between them to find the salience program which would be best for segmentation. Salience is the extent to which an object or area in an image pops out and catches the eye of the viewer. Previous work has shown that a breast tumor is more salient than other areas of the breast. Here we use a feature-based program to detect salience; the parts of the image that have the most-different features will be the most salient areas. In this feature-based program, the Gabor function is used to create a map from the original image, which will emphasize the parts of the picture that a simple cell in the mammalian visual system would see in the beginning stages of vision. In this step, the Gabor function will decompose the image into various orientation and frequency combinations. Next, the image is smoothed with a Gaussian filter, which has been proven to remove noise in MRI images. We use the Mahalanobis distance to find the pixels that are the most different from the mean of the image. Taking into account the Gabor and Mahalanobis functions, the image created emphasizes the parts of the image that a person would see in the early stages of vision and then finds the points of that image that are the most different from the other visible areas. These points that are the most different are the salient points, which are also the points most likely to be a breast tumor. Lastly, the Weibull function is applied to take into account the increased need in mammals for contrast when viewing an area that is either very dark or very bright. The result of the salience program is a grayscale image with the brighter parts of the image corresponding to the more salient parts of the image. In addition to our own program, the salience program which scored the highest will also be tested to see if using software which detects salience and tumors in mammograms would be as useful in detecting tumors in MRI. This salience program is created using a convolutional neural network (CNN). After the images are sent through the salience program, the segmentation is scored in its ability to segment images or it is input to a convolutional neural network (CNN) trained to create a segmentation map, which is also scored. A CNN is a computer algorithm that is biologically inspired in that the CNN can be described as a set of layers, each of which contain neurons or nodes, which “fire” or activate nodes in the next layer. Each neuron or node will be given a weight that will decide the extent to which each input node or neuron will affect the output segmentation. Three types of layers can be used in a CNN, but in our work only convolutional layers are used, making this program a fully convolutional network (FCN). Each convolutional layer is made up of filters whose weights will be updated with every back-propagation, transforming the data and allowing the FCN to create a more accurate segmentation. In this program, the salience program and FCN are combined to create a program that is more accurate at locating breast tumors than either program alone. To check that using the salience program is more accurate, a set of images are passed through both the salience programs and CNN program alone as well. In this paper, the highest accuracy for the CNN program alone is 0.92 and for the salience program combined with the CNN it is 0.98.