Faster RCNN Hyperparameter Selection for Breast Lesion Detection in 2D Ultrasound Images

Breast cancer is one of the leading forms of cancer. Breast lesion detection is a prerequisite for description of its characteristics and ultimately correct diagnosis of the lesion type. Compared to object detection in natural images, lesion detection in ultrasound images is a challenging task due to the inherent nature of the ultrasound images. This paper is concerned with adopting Faster Regions with Convolutional Neural Network (Faster RCNN) method for detecting breast lesions in 2D ultrasound images. Faster RCNN shows great promise in this application with very few misses of the actual lesion leading to high recall. However, the overall performance of this method suffers from high number of false positives. Therefore, we investigate different modelling hyperparameters of Faster RCNN to find an optimal configuration to improve the overall performance by essentially reducing false positives without significant increase in the number of misdetected lesions. Our empirical study using a total of 1183 images in three datasets shows that the optimal detection model outperformed original Faster RCNN due to significant reductions in false positives, resulting in 15% to 28% higher precision with only 3% to 11% drop in recall. Our optimal model also outperformed an existing breast lesion detection method by 3.4% in F1-score and provides better balance in precision and recall. Our study demonstrates that the optimal hyperparameter selection for Faster RCNN is a promising direction for breast lesion detection in ultrasound images.