You are on page 1of 1

Macro-scale Neuro Segmentation

Cori Harber, Lisa Meyer-Base, Erika Moreno, Hope Mumme, Heather Whittaker
Data Analytics for Engineers: Dr. A.J. Medford and Dr. Eva Dyer

Introduction Methods Results


Feature Extraction
Goal: Given a subpatch of a macroscale neuro-scan image, ● Feature Set 1: Entropy Mean-Variance-Sum, Shape Intensity M-V-S, Pixel
Intensity M-V-S, Histogram of Oriented Gradients (HOG) M-V-S
we can classify the subpatch by brain area.

Classes:

Feature Set #1 Feature Set #2

● Feature Set 2: Feature Set 1, and Average Alpha Class 0-1-2, Sum Alpha
Class 0-1-2, L2 Norm Alpha Class 0-1-2, L2 Norm Accuracy Class 0-1-2

Dataset: 92,160 total images, 128x128 are parent images


taken from patch, 64x64 images are rotated within the
subpatches

Test Accuracy for Feature Set #1


● The highest average testing accuracy was achieved prior to
dimensionality reduction at 97.6%.
● scaled PCA of the top 5 features came in first place with an average
accuracy across all classifiers of 85.21%.
● Kernel SVM using the rbf kernel did the best.
○ The best performing classifier in all five cases, achieving 97% test-set
accuracy.
Approach ● Performance for linear classifiers logistic regression and linear SVM
improved using feature set #2.
Feature Set #1 Feature Set #2
Dimensionality Reduction

3D Projection of Feature Set #1

Classification
● Train/Test split - 70 train, 30 test Conclusion
● k-Nearest Neighbors, kernel SVM, Decision Tree, linear SVM, ● Achieved 97% testing accuracy, with Kernel SVM and kNN
logistic regression. producing the best results, either without dimensionality
● The underlying distribution of the data is unknown, and the data is reduction or using scaled PCA with 5 components.
Baseline: A logistic regression model was trained using nearly linearly separable. Therefore, non-linear non-parametric ● The success of this work could potentially aid in automated
mean, variance, and sum of the entropy filter on the classifiers such as kNN, KSVM, DT are expected to do better than neuro-image segmentation for large-scale neuro-images
128x128 images. Training accuracy was 73.9% and testing linear parametric classifiers such as Linear SVM and LR. ○ alleviating manual annotation workloads for humans
accuracy was 74.7% ○ speeding analytical pipelines.
Kernel SVM Formula

You might also like