| Summary: | Breast cancer is a disease that affects millions of people. Several studies have identified breast density as an important risk factor for breast cancer. Thus, the evaluation of breast density is important for preventing breast cancer. Current commercially available ultrasound systems do not provide an estimation of breast density, and the evaluation of breast density is based on subjective visual observation of breast ultrasound images by radiologists; therefore, the accuracy of this evaluation is dependent on the skills of the radiologist, which may vary among radiologists. Several methods have been proposed to evaluate breast density in mammography and ultrasonography noting that there are several methods for mammographic evaluation but only a few for ultrasound evaluation. In this study, a set of breast ultrasound images was analyzed. Breast density was manually evaluated by two radiologists using this image set, including two distinct evaluations by the first radiologist in different periods. A quantitative and qualitative assessment was performed using semiautomatic and automatic algorithms with histogram thresholding algorithms and the Otsu method, resulting in six algorithms. An interval was defined for a quantitative analysis where the minimum value corresponds to the lowest value of the three radiologist observations, and the maximum value corresponds to the highest value of those observations. For the BDthr128 algorithm, 56% of the cases fall within the interval, whereas the value was 73% for the BDthrAuto algorithm; these findings show that the BDthrAuto algorithm has better performance than the former according to the radiologist evaluation of breast density. The application of an algorithm that isolates the mammary gland in BDthr128 and BDthrAuto resulting in BDCombo128 and BDComboAuto automatic algorithms is also described. The procedure used for the analysis of the breast density results was the same as that defined for the BDthr algorithms. After considering the range with the maximum and minimum for the observations of the same image, 28% of the values obtained by applying the algorithms were within the range for the BDCombo128 algorithm and 42% for the BDComboAuto showing that the automatic algorithm performs better according to the radiologist evaluations. Considering the three breast density observations for each image provided by radiologists and each breast density obtained with the four algorithms for each image, according to the qualitative BIRADS assessment, 3 hits were obtained for 33% of the 85 images using the BDthr128 algorithm and for 48% of the 85 images using the BDthrAuto algorithm. On the other hand, the BDthr128 algorithm achieved at least 2 hits with the radiologist observations in 69% of the images, whereas the BDthrAuto algorithm obtained 86% in the same situation. The BDCombo128 algorithm with 3 hits obtained 25% and the BDComboAuto algorithm obtained 47% in the same situation. With at least 2 hits, the BDCombo128 algorithm obtained 58%, and the BDComboAuto algorithm obtained 79%. For the application of the Otsu method, images with mammary nodules were not considered because based on the results obtained when using the previous algorithms, it was concluded that this type of image deserves special attention in future research. Thus and for the set of 82 breast ultrasound images, applying the BDthrOtsu semiautomatic algorithm, 65% of the images fall within the considered range, while for BDthrAuto this value is about 70%. Regarding automatic algorithms, BDComboAuto algorithm leads to 49% of images within the range, while the BDComboOtsu leads to 61%. For qualitative evaluation, with full coincidence with the radiologist observations, we obtained 46% of the values for the BDthrOtsu algorithm and the same value for the BDthrAuto. Thus, only in the quantitative assessment, the BDComboOtsu algorithm performs better than the BDComboAuto.
|
|---|