Automatic Lung Nodule Detection Using Deep Learning WEEK

Automatic Lung Nodule Detection Using Deep Learning WEEK 7: NODULE RADIUS ESTIMATION REU STUDENT: WINONA RICHEY GRADUATE STUDENT: NAJI KHOSRAVAN PROFESSOR: DR. BAGCI

Method Outline �Assumptions Region of inner nodule = high and homogenous intensity Region of outer nodule = lower average intensity �Threshold to give binary values THU – pixel value threshold – to determine if a pixel is part of the nodule PHU – percentage of positive area threshold – to determine if region is within the nodule �Edge detection Find transition from “inside nodule” to “outside nodule” Images: Ciompi, et al. 2015. Bag-of-Frequencies: A Descriptor of Pulmonary Nodules in Computed Tomography Images.

Radius Estimation: Function � Input: image patch � Output: estimated radius � Method Assume: Nodule Radius = edge between “inside” and “outside” Radius is sampled at ½ the patch radius using Bag of Frequencies Each point classified as inside or outside the nodule (pixel value threshold) � Radius is classified as inside or outside (percentage threshold) � If sample is inside nodule, your guess was too small � Endpoints: sample, edge If sample is outside nodule your guess was too big � Endpoints: center, sample A B Purple = sampled radius Blue = starting endpoint Yellow = ending endpoint A. Iteration 1: endpoints = 0, 20 Sampled radius, R = 10 Sampled intensities are inside the nodule Endpoints update to 10, 20 B. Iteration 2: endpoints = 10, 20 Sampled radius, R = 15 Sampled intensities are outside the nodule; endpoints become 10, 15

Method Parameters A B �THU – pixel value Purple = sampled radius Blue = starting endpoint Yellow = ending endpoint A. Iteration 1 7/8 intensities at R= 10 are above THU (7/8 >PHu are positive) radius is within nodule A. Iteration 2 3/8 intensities at R= 15 are above THU (3/8 <PHu are positive) radius is outside nodule is pixel in nodule? Based on HU of lung tissue �PHU – % of positive pixels is radius in nodule? �Number of iterations Accuracy = 20/2 iter Maximum iterations = 5 � Pixel rounding caps accuracy

Difficult Special Cases: Overestimation � Surrounded by other structures Non-Nodule high intensities cause an incorrect “inside nodule” ruling Radius can be drastically overestimated � Located on an edge Similar to above Less detrimental to estimation � Asymmetric in 3 D space Radius correctly identified for plane, overestimated for voxel � Uniform gradual intensity decline Every sampled intensity is just below the Threshold Overestimated Radius: surrounded nodule Radius: 2. 3 (red) Estimate: 20 (yellow) Overestimated Radius: Edge nodule Radius: 3. 5 (purple) Estimate: 7 (yellow) Overestimated Radius: Non-Symmetric nodule Radius: 4. 3 (purple) Estimate: 8 (yellow) Yellow, pink, and blue show the radius estimate at 100%, 50% and 25% respectively

Progress �Successful Implementation: Basic nodule radius estimation 2 D patch input �In progress: Improving accuracy by dealing with special cases

References �Ciompi, F. , Jacobs, C. , Scholten, E. T. , Wille, M. W. , de Jong, P. A. , Prokop, M. , & van Ginneken, B. (2015). Bag-of-Frequencies: A Descriptor of Pulmonary Nodules in Computed Tomography Images. IEEE Transactions On Medical Imaging, 34(4), 962 -973.