1. Automatic Lung Cancer Diagnosis from CT Scans (Week 2)
REU Student: Maria Jose Mosquera Chuquicusma
Graduate Student: Sarfaraz Hussein
Professor: Dr. Ulas Bagci

2. Tasks Accomplished
Unsupervised Learning Tutorial: Familiarization with unsupervised learning terminologies such as:
Bayes rule & Naïve Bayes (MAP & ML parameter estimates)
Latent variable models (Factor analysis, PCA, ICA, Mixture of Gaussians, K-means clustering)
EM Algorithm
Modelling time series and other structured data (SSMs, HMMs)
Intractability
Graph models (undirected graphs, factor graphs, direct graphs)
Etc.
Literatures:
1. Discriminative Unsupervised Feature Learning with Convolutional Neural Networks
2. Early Visual Concept Learning with Unsupervised Deep Learning
3. Towards Automatic Pulmonary Nodule Management in Lung Cancer Screening with Deep Learning
4. Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk
Scoring.
5. Multi-crop CNNs for lung nodule malignancy suspiciousness classification
6. Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks
7. Representation Learning: A Review and New Perspectives (in progress)
8. Why Does Unsupervised Pre-training Help Deep Learning? (in progress)
9. Unsupervised Learning Tutorial

3. 1: Discriminative Unsupervised Feature Learning with CNNs
Concentrates on concepts of invariance and ability to discriminate
CNN with no labeled data
Surrogate classes (sets)
Class created by taking a sample patch image and applying transformations it
Train network to discriminate among them .

4. 1: Discriminative Unsupervised Feature Learning with CNNs

5. 2: Early Visual Concept Learning with Unsupervised Deep Learning
Newborns are exposed to transformation of objects
Achieve zero-shot inference
Variational Autoencoder
Each latent factor (disentangled) learns to encode transformations

6. 3: Towards automatic pulmonary management in lung cancer screening with deep learning

7. 3: Towards automatic pulmonary management in lung cancer screening with deep learning
Multi-stream multi-scale architecture
CNNs process multiple triplets (sagittal, coronal, axial) of 2D views of a nodule at multiple scales
Gives probability for each class
Avoids nodule segmentation
Data augmentation
Observers vs. Computer

8. 3: Towards automatic pulmonary management in lung cancer screening with deep learning

9. 4: Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring
Convolutional Sparse Autoencoder (CSAE)
Task 1: Automated segmentation of breast density (MD)
Task 2: Characterize mammographic textural (MT) patterns
Convolutional layers are trained as autoencoders
Pre-training
Fine-tuning .

10. 4: Unsupervised Deep Learning Applied to Breast Density Segmentation and Mammographic Risk Scoring

11. 5: Multi-crop CNNs for Lung Nodule Malignancy Suspiciousness Classification
Multiple networks involves more computational costs
Multi-crop pooling layer (cropping feature maps): surrogates standard pooling
Malignancy suspiciousness classification using CT imaging
Problems: Nodule delineation due to morphology variation and capturing quantitative characteristics from nodule (heterogeneity)
Data augmentation

12. 5: Multi-crop CNNs for Lung Nodule Malignancy Suspiciousness Classification

13. 5: Multi-crop CNNs for Lung Nodule Malignancy Suspiciousness Classification

14. 6: Unsupervised Representation Learning with DCGANs
Deep Convolutional GANs
Use an all convolutional net (no pooling or fully connected layers)
Generator: Pooling replaced with fractional strided convolutions
Discriminator: Pooling replaced with strided convolutions (helps downsampling)
Batchnorm on both  helps with normalizing  efficient learning
ReLU activation in all layers (generator) except output (Tanh)  network learns quicker
LeakyReLU activation in all layers (discriminator)
Deduplication test in case of memorization
Generative models could model object attributes like scale, rotation, and position (less data required)

15. 6: Unsupervised Representation Learning with Deep Convolutional GANs