BoostCNN: Deep Learning AdaBoost-based Method for Easy and Difficult Nodule Classification in Ultrasound Images.
Authors
Affiliations (4)
Affiliations (4)
- Biomedical Engineering Department, Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, RJ, Brazil. Electronic address: [email protected].
- Universidade Federal do Ceará (UFC), Crateús, CE, Brazil.
- Biomedical Engineering Department, Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, RJ, Brazil; Núcleo de Investigación en Ciencias de la Motricidad Humana, Universidad Adventista de Chile, Camino a Las Mariposas, Chillán, Región de Ñuble, Chile.
- Biomedical Engineering Department, Universidade Federal do Rio de Janeiro (UFRJ), Rio de Janeiro, RJ, Brazil.
Abstract
Ultrasonography is used as a complementary imaging modality for breast cancer detection because it can detect cases missed by mammography. Nevertheless, global recommendations have not reached a consensus on using ultrasound as the primary screening tool. This is mainly due to the high number of false positives, which can lead to over-diagnosis, unnecessary treatment, surgical interventions and psychological stress. This paper aims to provide a novel AdaBoost-based ensemble method, called BoostCNN, that could help to reduce the rate of false positives as well as specialist false negatives in ultrasound images of breast cancer. We used the BUS-BRA dataset to train and test four state-of-the-art deep learning models as well as our proposed BoostCNN. For external validation, we then evaluated our methodologies on the BUSI and BrEaSt datasets. We obtained metrics for different numbers of models in the ensemble, highlighting the BoostCNN ensemble with nine models, which achieved an accuracy of 88.38 (95% CI: 86.93-89.81), sensitivity of 82.90 (95% CI: 79.86-85.86), specificity of 91.00 (95% CI: 89.37-92.48) and F<sub>1</sub>-score of 82.18 (95% CI: 79.79-84.43) on the BUS-BRA dataset. Furthermore, we also combined the ultrasonographer assessment and our BoostCNN prediction to evaluate joint performance, achieving an accuracy of 88.21 (95% CI: 86.72-89.65), sensitivity of 84.21 (95% CI: 81.22-87.09), specificity of 90.13 (95% CI: 88.43-91.67) and F<sub>1</sub>-score of 82.21 (95% CI: 79.77-84.50). These results demonstrate that our methods provide a potential tool to enhance specificity in breast cancer ultrasound classification. The Python implementation of our method is available in the following GitHub repository: https://github.com/PedroCrosara/BoostCNN.