1. Medical Physics Department, Faculty of Medicine, Urmia University of Medical Sciences 2. Department of Medical Physics and Imaging, Urmia University of Medical Sciences 3. Department of Radiology, Imam Khomeini Hospital, Urmia University of Medical Sciences
Abstract
Introduction
To explore diagnostic potential of computerizing texture analysis (ta) methods in discrimination of the normal, benign and malignant ovarian lesions by ct scan imaging.
Methods
The ovarian ct image database consists of 10 normal, 10 benign and 3 malignant which were reported by the radiologist and proven by clinical examination. the region of interests (rois) were defined within the lesion parts of the images. gray level intensity within a roi normalized by n1:default, n2:µ+/- 3σ, and n3:present intensity limited to 1%-99%. up to 270 features parameters extracted per roi per normalization schemes. among them, we selected subsets of ten best discriminating features based on two reduction methods: fisher (f) coefficient and or the probability of classification error plus average correlation coefficient (poe+acc). the selected features sets under standard and nonstandard states applied for ta with principle component analysis (pca) and linear discriminant analysis (lda). finally, the discrimination performance of the applied ta methods was evaluated by receiver operating characteristic (roc) curve analysis by calculation sensitivity, specificity, and area under the roc curve (az value).
Results
In differentiation of normal from benign ovarian lesions, pca in comparison with lda, represent excellent performance with a sensitivity 96.7%, specificity 80% and azvalue of 0.9. also in the differentiation of benign from malignant ovarian lesions, again pca represent high performance with a sensitivity 82.8%, specificity 96.5% and azvalue of 0.89.
Conclusion
Computerize texture analysis has high potential to promote radiologist's confidence for discrimination of ovarian lesions on ct scan images with no need other examination.
