1. Automatic Lung Cancer Diagnosis from CT Scans (Week 3)
REU Student: Maria Jose Mosquera Chuquicusma
Graduate Student: Sarfaraz Hussein
Professor: Dr. Ulas Bagci

2. Tasks Accomplished Programming/Other:
Familiarization with MATLAB and its Pre-Trained Neural Networks
Familiarization with ITK-SNAP Software
Literatures:
1. Representation Learning: A Review and New Perspectives
2. Why Does Unsupervised Pre-training Help Deep Learning?
3. Multiview Convolutional Neural Network for Lung Nodule Classification
4. Multi-stage Neural Networks with Single-sided Classifiers for False Positive Reduction and
its Evaluation using Lung X-ray CT Images
5. Pulmonary Nodule Classification with Deep Neural Networks on Computed Tomography
Images
6. An Information-Theoretic Framework for Fast and Robust Unsupervised Learning via
Neural Population Infomax

3. 1: Representation Learning: A Review and New Perspectives
General purpose priors for good representations
Deep Representations
PCA/ICA
Markov Random Fields
Bayesian Networks
Restricted Boltzman Machines
KL Divergence, Contrastive Divergence
Different types of Autoencoders (sparse, regularized, stacked, contracted)
Deep Belief Networks
Etc.

4. 2: Why Does Unsupervised Pre-training Help Deep Learning?
Pre-trained weights gathered through unsupervised learning are more beneficial then setting random weights
Better generalization
Training cost optimized  better regularization  less overfitting
Great for large layers and deeper networks; it can hurt smaller networks
Training and testing errors were lower
Act as a regularizer (maintained)
Advantage does not vanish as the number of samples increases
Variance reduction
Pre-train more layers  generalization

5. 3: Multiview Convolutional Neural Network for Lung Nodule Classification
Multiview CNN for classification of pulmonary nodules (malignant or benign)
Motivation: Multiple view provide different information
2 Tasks: binary and ternary classifications
Experiments performed for both tasks with 2 different input channel modes (1 and 7) and 5 different architecture with Batch Normalization
Batch Normalization used for faster learning and higher overall accuracy (but caused overfitting)
Deeper Network  separability of classes
Sensitivity, Specificity, ROC, t-SNE

6. 3: Multiview Convolutional Neural Network for Lung Nodule Classification

7. 3: Multiview Convolutional Neural Network for Lung Nodule Classification

8. 4: Multi-Stage Neural Networks with Single-Sided Classifiers
Cascaded multi-stage CNN with single-sided classifiers
Reduce the positives of lung nodule classification in CT scans
Irregular lesions in non-nodules create an imbalance
Inverse balanced dataset used
CNN stage filtering for non-nodules
Suspicious and obvious non-nodules
Threshold
Last CNN calculates probabilities

9. 4: Multi-Stage Neural Networks with Single-Sided Classifiers

10. 4: Multi-Stage Neural Networks with Single-Sided Classifiers

11. 5: Pulmonary Nodule Classification with Deep Neural Networks on Computed Tomography Images
Deep Convolutional Neural Network
Motivation: Diminish and ultimately reduce false positives and false negatives
Classify whether ROIs contain a nodule or non-nodule
2 Tasks: binary and ternary classifications
Experiments performed where tested CF and DD
Adjusted learning rate and momentum of weight updating
5 different trainings: training 4 gave most optimal results

12. 5: Pulmonary Nodule Classification with Deep Neural Networks on Computed Tomography Images

13. 6: An Information-Theoretic Framework for Fast and Robust Unsupervised Learning via Neural Population Infomax
Shannon’s Information Theory (closely related to MI and MAP)
Received signal ~ transmitted signal + noise
Monte Carlo (MC) sampling
Markov Chains
𝑋 →𝑌 → 𝑌 → 𝑌 →𝑅
Want to establish a relationship between 𝑋 and 𝑅

14. 6: An Information-Theoretic Framework for Fast and Robust Unsupervised Learning via Neural Population Infomax