Application of machine learning and deep learning methods on medical imaging aims to create systems that can help in the diagnosis of disease and the automation of analyzing medical images in order to facilitate treatment planning. Deep learning methods do well in image recognition, but medical images present unique challenges. The lack of large amounts of data, the image size, and the high class-imbalance in most datasets, makes training a machine learning model to recognize a particular pattern that is typically present only in case images a formidable task.
Experiments are conducted to classify breast cancer images as healthy or non-healthy, and to detect lesions in damaged brain MRI (Magnetic Resonance Imaging) scans. Random Forest, Logistic Regression and Support Vector Machine perform competitively in the classification experiments, but in general, deep neural networks beat all conventional methods. Gaussian Naïve Bayes (GNB) and the Lesion Identification with Neighborhood Data Analysis (LINDA) methods produce better lesion detection results than single path neural networks, but a multi-modal, multi-path deep neural network beats all other methods. The importance of pre-processing training data is also highlighted and demonstrated, especially for medical images, which require extensive preparation to improve classifier and detector performance. Only a more complex and deeper neural network combined with properly pre-processed data can produce the desired accuracy levels that can rival and maybe exceed those of human experts.
|
![[NJIT logo]](http://archives.njit.edu/vhlib/images/njit_logo.gif){kind=logo}
![[Van Houten Library logo]](images/vhl-top-image.jpg)
