Hierarchical Kernel Methods for Medical Image Analysis
Training machine learning systems on raw medical image data is believed to be hard due to the fact that the training images usually do not contain all possible variations. However, humans are good at learning a general concept behind a class of images from only a small number of examples. This ability is supportded by our visual system which is hierarchically structured and hierarchically extracts representations of retinal stimuli with increasing level of abstraction. It obtains more complex and abstract representations by composing simple representations extracted locally.
This master thesis project aims to implement and empirically examine a biologically inspired hierarchical architecture for breast cancer classification on mammographical images. The method was proposed by Smale et al. in the paper “Mathematics of the Neural Response” which provides a representation of images to reflect how humans see images. The experimental results show that for certain configurations the accuracy of 3-nearest neighbor classification on features extracted by the neural response method from mammographical images may override the accuracy achieved on the raw inputs. However, the average performance of this method does not live up to our expectations since the unsupervised learning rule does not work as good as expected from the literature. The experimental results also show that the breast coordinate system method proposed by Brandt et al. in the paper “An anatomically oriented breast coordinate system for mammogram analysis” is more suitable for extracting texture features. The breat coordinate system addresses the rotation variance in mammographical images and should be largely considered for improving the classification accuracy.
Supervisor: Christian Igel (DIKU)
Censor: Rasmus Larsen (DTU)