1. Automatic Lung Cancer Diagnosis from CT Scans (Week 4)
REU Student: Maria Jose Mosquera Chuquicusma
Graduate Student: Sarfaraz Hussein
Professor: Dr. Ulas Bagci

2. Background & Project Objective
Lung cancer is the first cause of mortality among all types of cancer
CT Scans
CAD Systems
Classify lung nodules as malignant or benign
using unsupervised learning methods
CAD Systems
Help radiologists with diagnosis: validation and blind spots.
They help save time.
Avoid misdiagnosis, which could lead to unnecessary surgeries, deaths, etc.

3. Previous Methods TumorNet 3D CNN Based-Multitask Learning TumorNet
Supervised
Multi-view CNN
3D-tensor (2D image patches)
Data augmentation (Gaussian, noise)
Gaussian Process Regression
1) Estimate testing labels
2) To get rid of inter-observer (radiologist) variation!
3D CNN Based Multi-Task Learning
Lung nodule dataset to fine-tune (transfer learning) a 3D CNN trained on Sport-1M dataset
3D CNN learns to discriminate features corresponding to 6 attributes
Graph regularized sparse representation: multi-task learning

4. What I Have Learned? Medical imaging concepts/terminology/software
Hospital Visit
Deep Learning concepts/terminology
Supervised
Unsupervised
MATLAB/Python
Supervised & unsupervised learning concepts/terminology
- Architectures that help unsupervised learning overall (Clustering, RBMs[Markov Chains], Autoencoders, GANs, etc.)
- Unsupervised pre-training to aid supervised learning
- Multi-task learning/ transfer learning (fine-tuning)
Unsupervised
- Representation learning

5. Literatures Review

6. Previous Approaches & Challenges
- Not enough labeled data
Lung nodules variations
Supervised learning architectures
3D CNNs vs. 2D CNNs
Multi-view/ Multi-scale
Fine-tuning pre-trained networks
- Unsupervised learning architectures
Representation learning
- Disentangled factors learning
- Invariance Properties
Autoencoders
0) Time consuming
1) False positive and false negative misdiagnose
2) Multi-view and multi-scale to aid in feature extraction for scans (volumetric)
3) Autoencoders: focus on getting all features and reconstructing images (not great)
4) Representation learning (stacked architectures, multi-task and transfer learning)
- Disentangled factor learning: transformation of data helps feature learning.
- Redundancy reduction
- Invariance properties of the learned representations

7. Tasks Accomplished (This Week)
Programming (MATLAB/Python):
Used Pre-Trained CNNs to Extract Features and Classify Lung Nodules from LIDC-IDRI dataset
Fined-tuned CNNs
GAN code
Literatures:
1. Generative Adversarial Nets
2. Lung Nodule Based on Computed Tomography Using Taxonomic Diversity Indexes and an SVM
3. Categorical Generative Adversarial Nets
4. Stacked Generative Adversarial Nets
LIDC-IDRI DataSet

8. Results (Supervised Classification)

9. Results (Supervised Classification)

10. Results (Supervised Classification)

11. Results (Supervised Classification)

12. Our Next Step…
- Study different GAN architectures and incorporate them
- Why GANs?
Unsupervised Learning
Have never been used
Generate instead of reconstruct
Better feature representation learning