امين ملك محمدي

Breast cancer is the most commonly diagnosed cancer among women. Breast screening is the most effective approach for reducing its mortality. To this end, Automated Breast Ultrasound (ABUS) is favorable due to its advantages over traditional mammography. Computer-aided systems for mass detection and segmentation are developed to assist the expert and find the symptoms in early stages. This study is composed of two sections; 1) A deep model which discriminates positive slices, and returns the approximate location of tumors as a heat map. 2) A saliency model which segments the tumors. In the first step, OBNLM is utilized to improve the quality of input slices. The architecture of the first model is inspired by the discriminator of well-known PatchGan model. Because of its patch based architecture, it can retrieve the location of tumors as heat map. The precision and recall of this model is increased by morphological operations. The effectiveness of this model is eva‎luated on a dataset of 60 volumes, belonging to 43 patients. According to FROC analysis, for two false positives, in region based and patient based analysis, the proposed model reaches the sensitivity of 82%, and 86% respectively. The saliency model uses a U-Net like architecture, which inputs a slice and its estimated saliency map to produce the final output. The first step of saliency map estimation is pixel clustering. To this end, K-means is utilized and resulting contours are filtered with respect to their size, contrast, and shape. This saliency map is combined with the first model’s output to produce the refined saliency map. Attention modules are deployed in every stage of encoder to combine the refined saliency map with network information and guide its attention. For the decoder, multiple connections are used to apply a deeper supervision and increase its recall. At the end of this network an SE module is deployed which increases the precision. With these modifications, the Dice similarity is increased by 6% compared to the base model, and reaches to 81%.

سرطان سينه , توده , آشكارسازي , بخش بندي

Breast cancer , Tumor , Detection , Segmentation

Amin Malekmohammadi

Dr. Mohsen Soryani