Automatic Lung Cancer Diagnosis from CT Scans (Week 1) REU Student: Maria Jose Mosquera Chuquicusma Graduate Student: Sarfaraz Hussein Professor: Dr. Ulas Bagci
Background Lung cancer is the first cause of mortality among all types of cancer. - CT (Computed Tomography) Scans CAD Systems (Computer Aided Diagnosis) 3D (sagittal, coronal, axial) Save time Avoid misdiagnosis, which could lead to unnecessary surgeries, death, etc. Validation and blind spots
Project Overview - Design a CAD system to automatically diagnose lung nodules without supervision - Techniques for extracting strong imaging features - Use neural network architectures Apply deep learning strategies Find features extracted from other architectures that could help with lung nodule classification
Problem Overview Extensive training of supervised models Time consuming Limited labeled data Regression accuracy not 100% yet
Tasks Accomplished - Studied neural networks and machine learning techniques - Familiarization with topic Background reading of previous work: 1. TumorNET: Lung Nodule Characterization Using Multi-View Convolutional Neural Network with Gaussian Process 2. Risk Stratification of Lung Nodules Using 3D CNN-Based Multi-task Learning 3. Characterization of Lung Nodule Malignancy using Hybrid Shape and Appearance Features - Understanding terminology Techniques used to extract features and classify tumor nodules Understanding how deep 3D CNNs work along with MTL and Transfer Learning And much more... .
Background Reading 1: TumorNet Architecture: Deep 3D CNN Input: 3D-tensor vector Training and testing: 10 fold cross validation Data Augmentation Dataset: LIDC- IDRI (1018 scans) Gaussian noise with random mean, Poisson, Salt & Pepper and Speckle noise GP to obtain malignancy score Joint distribution Conditional distribution Latent function Add Gaussian noise to latent function prevent overfitting in outputs Covariance between testing and training sets .
Background Reading 1: TumorNet
Background Reading 2: 3D CNN-Based Multi-task Learning - 3D CNN trained on Sports-1M dataset - Dataset: LIDC-IDRI, 10 fold cross validation - Transfer learning fine-tune 3D CNNs using labels for malignancy and six attributes. 1 CNN per attribute L1 norm - Multi-task learning Graph Regularized Sparse Representation Trace norm Correlation of different tasks: Coefficient Matrix W Structure Matrix Correlation Matrix Binary Matrix Scoring function - Malignancy score
Background Reading 2: 3D CNN-Based Multi-task Learning
Background Reading 3: Hybrid Shape and Appearance Features - Spherical Harmonics Computation and pre-trained DCNN - 10 fold cross validation - Two inputs: Radiologists’ binary nodule segmentations mesh representation SH functions Vectors Local CT image Combine three orthogonal local patches (applied PCA on binary segmentation data - Extracted representation from SH method provides benefits (rotate, scale, transform, easier to compute) - RF Classification Synthesize DCNN and SH features - Malignancy evaluation
Background Reading 3: Hybrid Shape and Appearance Features
Background Reading 3: Hybrid Shape and Appearance Features