Shape Descriptors and Fourier Descriptors for Image Retrieval

Shape Descriptors and Fourier Descriptors for Image Retrieval Med. IX Program REU Presentation by: Marwa Nur Muhammad Advisors: Dr Daniela Raicu and Dr Jacob Furst

Motivation n Increase the accuracy of medical diagnostics Give radiologists comparable and similar images Make the diagnostics less subjective

Goals n Look at shape and boundary features, as opposed to texture features by: q q n Developing Shape Descriptors Developing Fourier Descriptors Add to BRISC

Previous Work Done n n BRISC (Content-based image retrieval system) Texture based retrieval systems implemented q q q Gabor Markov Co-occurrence

Shape versus Fourier n n n Shape captures the region information Fourier captures boundary information Fourier is invariant to: q q q Translation Rotation Scaling

Shape Retrieval n n n Calculate the 8 shape features on the images in LIDC Find the Euclidean Distance (ED) between the features of the query image and the features of the other images Retrieve the best images (images with shortest ED) Calculate precision by: The image is considered to be relevant if it is from the same nodule as the query image

The 8 Shape Features n n n n Circularity Roughness Perimeter Major axis length Elongation Convex Perimeter Compactness Eccentricity Minor axis length Convex hull Solidity Extent Radial Distance Standard Deviation

Fourier Descriptors n n Uses contour information Different signatures that can be used are q q n n Complex coordinates Centroid distance Curvature signature Cumulative angular function Centroid distance method used Different images have different number of contour points marked by the radiologists q Ranged from 9 to 311

Radiologist rating 1 Nodule 1 Radiologist rating 2 Case/ Patient Radiologist rating 1 Slice 1 Contour 1 Slice 2 Contour 2 Slice 3 Contour 3 Slice 1 Contour 1 Slice 2 Contour 2 Nodule 2 Radiologist rating 2

Examples

Do Images with Larger Area have more Contour points?

Steps n n n Count boundary points Equal arc length sampling Calculate centroid distance n Divide magnitude of first half of FDs by DC component n Compute Fourier transform values

How many boundary points? 8 points 32 points 64 points

Equal Arc Length Sampling n 8 to 31 points 8 equally spaced contour points n 32 to 63 points 32 equally spaced contour points 64+ points 64 equally spaced contour points n

Centroid Distance n X-coordinate n Y-coordinate

Precision n n Calculating Euclidean distance for images with the same number of contour points Calculating Hausdorff distance for images with different number of contour points

Results for Shape Descriptors

Results for Fourier Descriptors

Results for Fourier Descriptors

Prediction model for Shape Descriptors Nodule Important charact shape eristics features Lobulation elongation, roughnes s Malignanc y Margin Sphericity - Spiculation radial distance SD Subtlety -

Prediction model for Fourier Descriptors n Characteristics used q q Lobulation Margin Sphericity Spiculation

Most Important Fourier Descriptors (for nodules with 64 boundary points)

Classification Matrix for Lobulation (for nodules with 64 boundary points) Tree depth = 50; Parent node = 5; Child node = 2

Classification Matrix for Margin (for nodules with 64 boundary points) Tree depth = 50; Parent node = 5; Child node = 2

Classification Matrix for Sphericity (for nodules with 64 boundary points) Tree depth = 50; Parent node = 5; Child node = 2

Classification Matrix for Spiculation (for nodules with 64 boundary points) Tree depth = 50; Parent node = 5; Child node = 2

Conclusions n n Decision trees show that shape might not be too effective for retrieving similar nodules Decision trees for Fourier descriptors show relatively low error rates q Good results can be expected from Fourier transform

Future Work n n Use Hausdorff distances as well Centroid distance might not have given best results Use complex coordinates, cumulative angles and curvature functions to see which one produces best results Try other methods such as taking moments

Acknowledgments n n Ekarin Varutbangkul Aaron Mintz CTI, De. Paul University Med. IX REU Program

References 1. M. Lam, T. Disney, M. Pham, D. Daniela Raicu, J. Furst, R. Susomboon. Content-based image retrieval for pulmonary computed tomography nodule images. Medical Imaging 2007: PACS and Imaging Informatics, Proceedings of the SPIE, Volume 6516, March 2007. 2. D. Raicu, E. Varutbangkul, J. G. Cisneros, J. D. Furst, D. S. Channin, S. G. Armato III. Semantics and Image Content Integration for Pulmonary Nodule Interpretation in Thoracic Computer Tomography. SPIE--The International Society for Optical Engineering. March 3 2007. 3. D. Zhang, G. Lu. A Comparative Study on Shape Retrieval Using Fourier Descriptors with Different Shape Signatures. The 5 th Asian Conference on Computer Vision. 23 -25 th January 2002.