2019 · 99 citations · 16 references
ComputingComputer ScienceRadiologyMachine LearningUltrasonographyUltrasoundDeep LearningMedical Image AnalysisMedical ImagingOverfittingTransfer LearningUltrasound ImagingBiomedical ApplicationsConvolutional Neural NetworksDeep Learning In RadiologyDeep Learning-based ClassificationDeep Learning-based DetectionDeep Learning-based Lesion ClassificationMalignant Breast TumorsMalignant Diagnosis
Computer-aided detection of malignant breast tumors in ultrasound images has been receiving growing attention. In this paper, we propose a deep learning methodology to tackle this problem. The training data, which contains several hundred images of benign and malignant cases, was used to train a deep convolutional neural network (CNN). Three training approaches are proposed: a baseline approach where the CNN architecture is trained from scratch, a transfer-learning approach where the pre-trained VGG16 CNN architecture is further trained with the ultrasound images, and a fine-tuned learning approach where the deep learning parameters are fine-tuned to overcome overfitting. The experimental results demonstrate that the fine-tuned model had the best performance (0.97 accuracy, 0.98 AUC), with pre-training on US images. Creating pre-trained models using medical imaging data would certainly improve deep learning outcomes in biomedical applications.
16
Dropout: a simple way to prevent neural networks from overfitting
Nitish Srivastava, Geoffrey E. Hinton, Alex Krizhevsky et al. · 2014 · 34.3K citations