Deep Learning to Improve Breast Cancer Detection on Screening Mammography

Abstract

The speedy development of deep learning, a family of machine learning techniques, has spurred abundant interest in its application to medical imaging issues. Here, we have a tendency to develop a deep learning algorithmic program which will accurately sight carcinoma on screening mammograms victimization AN “end-to-end” coaching approach that with efficiency leverages coaching datasets with either complete clinical annotation or solely the cancer standing (label) of the complete image. during this approach, lesion annotations square measure needed solely within the initial coaching stage, and future stages need solely image-level labels, eliminating the reliance on seldom offered lesion annotations. our all-convolutional network technique for classifying screening mammograms earned glorious performance compared with previous strategies. On AN freelance take a look at set of digitized film mammograms from the Digital info for Screening diagnostic technique (CBIS-DDSM), the most effective single model achieved a per-image United Self-Défense Force of Colombia of zero.88, and four-model averaging improved the United Self-Défense Force of Colombia to zero.91 (sensitivity: eighty-six.1%, specificity: eighty.1%). On AN freelance take a look at set of full-field digital diagnostic technique (FFDM) pictures from the IN-breast info, the most effective single model achieved a per-image United Self-Défense Force of Colombia of zero.95, and four-model averaging improved the United Self-Défense Force of Colombia to zero.98 (sensitivity: eighty-six.7%, specificity: ninety-six.1%). we have a tendency to additionally demonstrate that a full image classifier trained victimization our end to-end approach on the CBIS-DDSM digitized film mammograms may be transferred to IN breast FFDM pictures victimization solely a set of the IN-breast information for fine-tuning and while not any reliance on the supply of lesion annotations. These findings show that automatic deep learning strategies may be without delay trained to realize high accuracy on heterogeneous diagnostic technique platforms, and hold tremendous promise for up clinical tools to scale back false positive and false negative screening diagnostic technique results.