Ingrid Sluimer, PRESENTER: Nothing to Disclose

Abstract: HTML Purpose: The extraordinary detail in HRCT of the lung makes it a difficult modality to assess, calling for the development of computer-aided diagnosis (CAD) software to assist the differential diagnosis. A computer system was developed to classify regions of interest (ROIs) into one of 5 categories, based on appearance. The categories are normal, linear and reticular, nodular, high opacities, and emphysema. Methods and Materials: 167 scans (of different subjects) were collected, representative of what is encountered in daily clinical practice. Scans with movement artifacts or severe streaks were excluded. Circular ROIs were extracted with an 80 pixel diameter (physical dimension ranging from 26 mm to 66 mm). Slice thickness was either 1.0 or 1.5 mm. The final dataset consisted of 384 ROIs taken from 116 scans. Classification of an expert chest radiologist was used as gold standard. This resulted in the following amounts of ROIs per category: normal (90), linear and reticular (90), nodular (52), high opacities (55), and emphysema (97). A filter bank including filters based on the Gaussian and its derivatives was employed to calculate a feature vector for each ROI. A k-nearest neighbor (kNN) classifier was used to classify the feature vectors. Results: Performance was tested in leave-one-patient-out experiments. Training involved normalization of the feature vectors, automatic feature selection and training of the kNN-classifier. Out of 384 samples, 199 were correctly classified by the computer program (average accuracy 52%). Accuracies per category were: normal (58%), linear and reticular (58%), nodular (17%), high opacities (46%), and emphysema (63%). Conclusion: The automated system was able to place more than half of the ROIs in our dataset into the correct category. Given the large amount of subtle cases in our dataset, this is a promising result. Performance is expected to improve with increasing size of the data set.       Questions about this event email: I.C.Sluimer@azu.nl

Sluimer, I, Texture Analysis for Automated Classification of Lung HRCT Images.  Radiological Society of North America 2003 Scientific Assembly and Annual Meeting, November 30 - December 5, 2003 ,Chicago IL. http://archive.rsna.org/2003/3106331.html