Computer Vision Brain Segmentation

Project Definition, Audience and Scope:

Currently, Convolutional Neural Networks (CNN) have become recognized as a new efficient method in classification. CNNs have proven to be more effective at image classification than humans, which has established the use of deep learning in brain imaging. Thus CNN has rapidly become the choice in improving brain diagnosis, treatment and characterization, which benefits from accurate segmentation.

For example, the ability to segment and visualize the thalamus would allow us to determine whether trauma or disfigurement of certain areas of the thalamus would indicate the possibility of developing certain disorders or syndromes. One example of spatial specific injury that leads to disorder is trauma to the Anterior thalamic nucleus, which would lead to the development of anterograde amnesia. While current methods require us to segment and analyze by hand, our goal is to optimize this process for faster and more accurate diagnosis.

Our Solution:

Currently, Convolutional Neural Networks (CNN) have become recognized as a new efficient method in classification. CNNs have proven to be more effective at image classification than humans, which has established the use of deep learning in brain imaging. Thus CNN has rapidly become the choice in improving brain diagnosis, treatment and characterization, which benefits from accurate segmentation.

Product:

Our CNN was trained on 10 images and 8 for testing purposes. After training on 10 images, we obtained a very clear and fluid segmentation and used an ADAM optimization of 0.0001 in our CNN for highest accuracy. We utilized 5 epochs because less epochs displayed less feature data from the input image while more epochs had negligible improvement. We were careful not to overfit the data, which results from training on the same data set repeatedly.

Additionally, we varied our CNN by changing the kernel size. Larger kernel size meant more required memory for weights and more computation time for back-propagation. Smaller kernel size meant a drop in accuracy. Thus we chose a 3x3x3 kernel, which straddle the two extremes. Our CNN performed poorly when we experimented with training on all 43 individual structures. Our filter size provided limited feature detection, which can be an area for improvement. We may also improve network performance by increasing the number of convolutional layers to deepen the network, which would boost performance by learning more complex, non-linear behaviors.

Evaluation:

we had decided to experiment with a convolutional neural network to accurately segment subcortical structures in brain MRI image scans. Our tests indicate that our proposed method can be an effective and efficient approach at brain segmentation for diagnosis, however, we must understand that there are limitations: currently our model cannot detect more structures as it had trouble learning more than our seven proposed classes. Also, due to system and memory limitations, we couldn’t get more accurate results. However, with better infrastructure, and a deeper neural network, segmentation can become more accurate and better optimized.