B-mode 부분 영상을 이용한 커리큘럼 학습 기반 유방 초음파 영상 분류

Abstract

While ultrasound examination is widely used for breast tumor diagnosis due to its non-invasive and eco- nomic advantages, classification accuracy is inherently affected by ultrasound image quality and diagnostic subjec- tivity. Recent advancements in deep learning have introduced automated analysis methods for classification of medical ultrasound images, achieving high accuracy. However, the performance of these methods critically depends on the availability of large-scale datasets. This paper proposes the classification method using curriculum learning with B-mode partial images to achieve high classification performance with limited datasets. By starting learning from small partial images centered on tumor regions defined by mask images and progressively expanding to include wider contextual areas, the proposed method enables the model to progressively learn ultrasound image characteristics. To validate the proposed method, we conducted experimental evaluation using 200 breast ultrasound images from the OASBUD dataset. Experimental results demonstrated that curriculum learning significantly improved classifi- cation accuracy compared to original image learning and data augmentation