Premium
WE‐C‐332‐08: Preliminary Analysis of Morphological Features From T1 and T2 MR Images in the Diagnosis of Breast Cancer
Author(s) -
Hipp E,
Bhooshan N,
Giger ML,
Arkani S,
Li H,
Newstead G
Publication year - 2008
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.2962722
Subject(s) - pattern recognition (psychology) , artificial intelligence , linear discriminant analysis , receiver operating characteristic , correlation , feature selection , feature (linguistics) , breast cancer , feature extraction , mathematics , computer science , medicine , cancer , statistics , linguistics , philosophy , geometry
Purpose: To demonstrate the potential of computer‐extracted morphological features of lesions on breast MRI using T1 and T2 images to distinguish between malignant and benign breast lesions. Method and Materials: A pilot database of 36 breast lesions: 18 malignant and 18 benign masses as determined by biopsy, was compiled. Images were acquired as coronal T1‐weighted spoiled gradient echo sequence images and axial T2‐weighted fast spin echo images. Lesions were segmented using fuzzy c‐means clustering and features were automatically extracted. Classification performance was investigated for the T2‐extracted features using receiver operating characteristic analysis (ROC). Employing stepwise selection, linear discriminant (LDA) round‐robin, and ROC analysis, merged‐feature performance was also assessed. Statistical Z‐tests were performed to determine significance. Additionally, correlations between feature on both the T1 and T2 image were investigated. Results: Classification performance using three T2‐extracted features (Correlation, Energy, and Irregularity) yielded an AUC value of 0.69, which was found to be statistically significance compared to an AUC of 0.5 (p = 0.05). While similar performance was found for morphological lesion features from T1 images, only moderate correlation was observed between T1 and T2 features, with contrast and entropy features demonstrating a moderate positive correlation. Conclusion: Classification performance of T2‐extracted feature combinations from this preliminary dataset is significant for distinguishing between malignant and benign. Additionally, these preliminary results suggest further study of the extent to which T1‐ and T2‐ extracted feature correlation can be exploited. Further investigation is necessary to assess the effectiveness of computer‐extracted morphological features in characterizing benign and malignant lesions. Conflict of Interest: Research supported in part by NIH. Some authors receive royalties, research funding, and/or are stock holders in Hologic.