Deep Learning in Medical Image Analysis: Alzheimer’s Disease Classification & Whole Fetal Brain Segmentation.

Abstract

During the last couple years there has been a growing interest in Deep Learning (DL), thanks in part to an increasing availability of computational power and of massive labeled datasets. This work presents two different Deep Learning applications for medical image analysis related tasks.

The first application is the classification of Alzheimer Disease (AD) using MRI. AD is the most common form of dementia. However, it’s difficult to diagnose. During this work three different DL models are proposed to classify subjects into either AD or Normal Control (NC) groups. The models proposed are a custom CNN model, vgg16 + Transfer Learning and vgg16 + Fine Tuning. The vgg16 based methods make use of pre-learned weights trained on the ImageNet dataset. All of the models used as an input a single 2D MRI slice. This helped reduced model complexity and training times. After performing a 10 times repeated 5-Fold cross validation it was discovered that the best model was vgg16 with Fine Tuning. It reached an accuracy of 80% and an AUC of 0.86. The obtained results proved that the use of weights learned from different domains could be useful in medical image applications. The main contributions of this section were the use of a single MRI slice to classify patients, the model validation technique and the use of transfer learning.

The second application is the fully automatization of the Whole Fetal Brain Segmentation in maternal MRI scans. For this U-Net [41] and other segmentation networks were compared. The U-Net model was also modified to include two types of attention components. Squeeze and Excitation module [16] and Attention Gates [39]. To avoid overfitting the models were trained using rigorous regularization (channel wise dropout, weight decay and data augmentation). They were also trained using different loss functions, including a Hybrid (BCE+Dice) loss function. The results of the 10-fold cross validation experiments showed statistically that no model was able to outperform the vanilla U-Net (DSC of 0.93 and HD 7.59mm). The proposed Attention Gated U-Net trained with the Hybrid loss reached a 0.95 DSC and 5.03mm HD (no significant difference). However, results on challenging data show a better performance for the Attention Gated models. This led to the conclusion that the possible advantage of using attention gates outweighed the cost of adding the attention module. The results were also improved by the use of a post processing step based on traditional computer vision techniques. The main contributions to this problem were the use of 10-fold cross validation, the statistical comparison between models, the use of the Hausdorff distance metric, the post processing step and the addition of attention components.

This work also provides an overview of other deep learning applications also used within a medical setting. The techniques mentioned include image labeling, autoencoders and GANs for data augmentation. It also provides an overview of the current challenges and limitations DL faces and the most common way of mitigating them. Thanks to the research conducted during this thesis it can be concluded that the use of DL can be used in a medical setting focused on the analysis of images. Particularly on classification and segmentation tasks. However, certain challenges such as having a large enough data set and availability of the required hardware to carry out research must be considered before attempting to solve a problem with DL.