Résumé: Manual segmentation of breast lesions is a tedious and time-consuming task. The existing state-of-the-art studies are evaluated on single modalities, i.e., either on mammogram or ultrasound thereby having limited clinical application. The limitations stated above motivated us to develop a Computer-Aided Mammogram Segmentation (CAMS) system and Computer-Aided Ultrasound Segmentation (CAUS) by using a customized pre-trained AlexNet network to perform semi-automated segmentation of breast lesions in dual-modality. A new real-time dual-modality data of mammogram and ultrasound are developed for conducting this study. Data augmentation is applied to improve diversity of the data. The last layer of the pre-trained network is modified to perform pixel-wise classification of the input test image. The output binary image is obtained using the color map of each test image. Various performance measures and statistical correlation values of paired T test are used for evaluation of the proposed model. The proposed model achieves a Jaccard index of 0.53 for mammogram and 0.63 for ultrasound, respectively. Further, a Dice similarity coefficient of 0.66 for mammogram and 0.76 for ultrasound is achieved using the proposed CAMS and CAUS systems, respectively. Area under curve of the proposed CAUS system is found to be 0.81 which is very close to that of the CAMS, i.e., 0.82. It is concluded that the CAUS system can be useful in breast lesion detection along with CAMS in the routine clinical scenario. Ultrasound being a painless, ionizing radiation-free, and low-cost technology can be useful for breast cancer screening over traditional mammograms, particularly in low settings.

Langue: Anglais