Volumetric Medical Classification using Deep Learning: A comparative study on classifying Alzheimer's disease using Convolutional Neural Networks

Thesis / Dissertation


Permanent link to this Item
Journal Title
Link to Journal
Journal ISSN
Volume Title
This work sets about designing and implementing a number of deep-learning models capable of identifying Alzheimer's disease from MRI brain scans. A common problem with detecting the disease is the difficulty in doing so before outward mental symptoms have begun to show. Therefore, the models attempt to classify both mild and severe cases. The experimental process proves that a problem involving volumetric medical images benefits from the usage of 3D model architecture over traditional 2D architecture. In doing so, however, it is revealed that the 2D models do ultimately perform only slightly below the 3D model. Thus, the 2D approaches hold merit for potential usage, should a 2D planar approach be desired. The paper presents a total of three models. The first is a 3D CNN model, which performs the best in all regards, with a mean accuracy of 81.3%. It is treated as the optimal means of detecting Alzheimer's. The second is a 2D CNN model which uses separate 2D convolution layers to independently train and combine 2D slices across the depth axis. This approach produces a model that only slightly under-performs compared to the 3D model (80% accuracy). The third and final model is a novel design in which a set of models are each trained on a single unique 2D slice of the volume, across a carefully chosen range of slices deemed to contain the most favourable feature data. The model set is then used in unison to make predictions which are then aggregated using a weighted ensemble-voter to produce a final prediction score. This final design scored between the prior two models (80.6%), and establishes itself as a promising model capable of operating on a fraction of the data. Analysis of the models' activation gradients was conducted to confirm that 2D models are able to train well on isolated 2D slices, but struggle to process the space between these slices. Additionally, the work examines and rates the effectiveness of several optional variables in the overall CNN model design, specifically in the context of training on brain scans. A variety of pixel rescaling functions were found to have a noticeable positive impact on overall model performance. Regularization, as well as augmentation in the form of rotation / elastic deformation, also yielded similar improvements on such models, and are thus universally recommended as considerations for any works attempting to solve a similar classification problem. With all this in mind, a final conclusion is made that machine learning and deep learning are promising tools in the medical field for assessing and diagnosing using raw brain scans. For additional reference, the code repository for generating and processing the models is available for viewing. An alternate branch, containing the code used to produce the gradient activation maps, has also been included.