Using Two Layers of BI-RADS Features for Locating Similar Calcification Lesions

With the rapid advancement of imaging technology, an inordinate amount of digital medical images have been generated in hospitals and medical institutions. To exploit those medical images in an effort to aid the diagnoses and research, content-based image retrieval systems are required to effectively access the medical image databases. This study presents a content-based image retrieval system which enables medical professionals to locate calcification lesions that are pathologically similar to a given example. More importantly, the type and distribution features of calcification lesions are extracted to represent the characteristics of mammographic lesions according to the Breast Imaging Reporting and Data System (BI-RADS), which is widely utilized by radiologists to describe mammographic lesions. In performance evaluation, a mammogram dataset was used to assess the effectiveness of the extracted type and calcification features. Our experimental results demonstrated that when the retrieval system only compares the calcification type and/or the calcification distribution characteristic, the pleomorphic type presents a higher precision-recall curve than the other three varieties of calcification types, and the cluster distribution performs best among the three lesion distributions. When the retrieval system takes both type and distribution characteristics into consideration, the pleomorphic and clustered class shows the best performance amongst all the calcification lesions.