Breast Tumors Delineation for Mammogram Investigations Exploiting k-Means Algorithm and Associated Image Processing
Mohamed Hisham Aref, Ayman Ahmed Nassar, Amr A. Sharawi, Mohamed A. Abbass, Ahmed M. Awadallah, Mohamed Rabie, Ayman Mohammed Farag, Sara Abd El-Ghaffar, Yasser H. El-Sharkawy · 2022
Breast cancer early detection is critical for life-saving and fast healing. Radiologists can misdiagnose many cases due to the modality’s low accuracy, or lack of experience. We aim to customized a computer-assist detection (CAD) system to assist the radiologists in the early discovery of the malignants in the mammograms. This study involved 150 mammograms for patients with various breast masses thru two mammographies: (GE Sys., USA) (N = 80) and MAMMOMAT Revelation (Siemens, Germany) (N = 70)). Normal cases (N = 30) were used in the system as a control group to evaluate the specificity of false positive (FP) and true negative (TN) results. Mammogram images were processed by a custom algorithm (Normalization, moving average filtering, and K-means clustering) to automatically differentiate between normal and malignant regions. Although, experimental results were validated by the "Lunit inc. online software" to compare the tumor detection accuracy. However, all cases were re-checked and assessed by two qualified radiologist readers. The system presents reasonable outcomes concerning CAD for breast tumors regarding the qualified radiologist, where out of the 150 patients, twelve were missed by the custom CAD algorithm and five were missed by at least one of the radiologists, although the CAD sensitivity and specificity were 92.5% and 90%, respectively. The presented framework improves breast cancer detection from the mammogram investigation regarding two qualified radiologist readers with an overall accuracy of 92%. The presented CAD system could assist the radiologist to avoid missing the recognition of breast tumor cases, in addition to offering the possibility to improve careful tumor resection during surgery.