Classification of Mammogram Analysis Using BL-CNN
Hichem Razgallah, Jiyun Li · 2024
One of the methods currently available for breast cancer screening is mammography, which can be used to detect lesions related to breast cancer. It is known that tumors can occur in any part of the breast, and that a single breast may have multiple lesion instances. It can be challenging to identify multiple suspicious locations and classify them appropriately in the meantime. To tackle this problem, we proposed BiLinear-CNN (BL-CNN in short) in which a Mask R-CNN model is used to detect and extract the lesion instances and a bilinear pooling model to classify them as benign and malignant. BL-CNN can identify the suspicious regions of a single breast with a confidence score and classify them as either benign or malignant in the meantime. The Mask R-CNN part of BL-CNN is pretrained with COCO Dataset for lesion instance detection, and the Bilinear Pooling part is pretrained with ImageNet dataset for classification. The models are fine tuned with breast cancer mammography dataset. Experiments on two breast cancer datasets CBIS-DDSM and Inbreast have shown the effectiveness of our model. This model achieved 90% AUC and 95.58% overall accuracy performance over the SegNet segmentation model and U-Net CNN encode and decoder architecture.