Automatic Lung Nodule Detection Using Deep Learning WEEK

Automatic Lung Nodule Detection Using Deep Learning WEEK 6: FINE TUNING REU STUDENT: WINONA RICHEY GRADUATE STUDENT: NAJI KHOSRAVAN PROFESSOR: DR. BAGCI

Week Summary �Tested Handcrafted features Ran on all nodules samples Continued visualizing results Made slight modifications to optimize feature vectors �Feature Selection Method Review Feature Selection with Annealing

Feature Selection with Annealing: Overview �Combining the regularization technique and the sequential algorithm design �Shrinkage estimation Gradually removing less relevant variables based on a criterion and a schedule �Annealing to lessen greediness Probabilistic technique to estimate global maximum �Problem size drops throughout iterations Ideal for big data learning

Important Improvements �Computational Efficiency Problem size decreases with each iteration Easily distributed over a grid of processors �Theoretical Guarantees Convergence, variable selection, and parameter consistency �Prediction power �Feature selection accuracy No undesired bias introduced (unlike penalized loss) �Can be applied to any loss function �Parameter, k = # of relevant features Intuitive, easy to specify Robust

Algorithm Outline Initialize parameter vector, β = 0 2. For given # of iterations 3. Update parameter, β, to minimize loss, L(β) by gradient descent 1. 4. Variable selection: 1. 2. Keeps only highest coefficient magnitudes, |βj| Renumber coefficients, 1, …, Me in N

Next Steps �Nodule radius estimation for input into handcrafted features �Feature selection implementation

References �Barbu, A. , She, Y. , Ding, L. , & Gramajo, G. (2013). Feature Selection with Annealing for Computer Vision and Big Data Learning.