Using Forestnets for Partial Fine-Tuning Prior to Breast Cancer Detection in Ultrasounds

Dalia Rodríguez-Salas, Mathias Öttl, Mathias Seuret, Kai Packhäuser, Andreas Maier

Abstract: Ultrasound imaging is the most used technique for first investigating suspicions of breast lesions. In recent years, many deep learning techniques have been developed to analyze ultrasound imagery automatically. In this paper, we present results obtained on breast lesion detection using FasterRCNN with a ResNet18 as the backbone. We pre-trained the backbone on classifying benign and malignant lesions using an extremely sparse classification head called ForestNet instead of the usual fully-connected head. Results on classification showed that the inherent sparsity of ForestNet led to reduced overfitting and kept untouched roughly 7% of the millions of parameters in the ResNet18 backbone. Our experiments were run on two different datasets. On classification, our modified ResNet-18 architecture classified correctly 85.9% and 91.3%, while the usual architecture achieved 75.6% and 80.6%, respectively. Regarding detection, the average precision on the two datasets rises from 80.4% and 79.2% to respectively 87.3% and 87.0% using our approach.