AugNet: A Deep Convolutional Neural Network for Classifying Natural versus Augmented Breasts
Emily A. Borsting, MD, Robert D. DeSimone, B.S..
Department of Plastic Surgery, UC Irvine, Orange, CA, USA.
PURPOSE -
Recent advances in deep learning and artificial intelligence have been transformative in the fields of computer vision and natural language processing, as well as in other domains like radiology, genomics, and pharmaceutical research. Herein we present to our knowledge the first application of deep learning to the field of plastic surgery. Our goal was to demonstrate a concrete example of a deep neural network performing at or beyond the level of a human expert.
METHODS -
A convolutional neural network (AugNet) was developed to classify breast augmentation images using only pixels and augmentation status labels (“before”/“after”) as inputs. AugNet was trained using a dataset of 50,000 before and after photos which were collected from the public web. The network's performance was tested against 12 plastic surgery residents on 4,435 previously unseen images.
RESULTS -
AugNet was accurate on 97% of the test-set images, outperforming 11 of 12 plastic surgery residents.
| Rater | Recall | Precision | F1 Score | Accuracy |
| Resident 1 | 0.9618 | 0.9844 | 0.973 | 0.9721 |
| AugNet | 0.9689 | 0.972 | 0.9704 | 0.9705 |
| Resident 2 | 0.9706 | 0.9706 | 0.9706 | 0.9697 |
| Resident 5 | 0.9421 | 0.9813 | 0.9613 | 0.9629 |
| Resident 3 | 0.9535 | 0.9762 | 0.9647 | 0.9603 |
| Resident 4 | 0.9545 | 0.9692 | 0.9618 | 0.9603 |
| Resident 7 | 0.9592 | 0.9592 | 0.9592 | 0.9588 |
| Resident 8 | 0.9733 | 0.9406 | 0.9566 | 0.9565 |
| Resident 6 | 0.954 | 0.9651 | 0.9595 | 0.9542 |
| Resident 9 | 0.9438 | 0.9438 | 0.9438 | 0.9359 |
| Resident 10 | 0.8864 | 0.975 | 0.9286 | 0.9286 |
| Resident 11 | 0.9167 | 0.9565 | 0.9362 | 0.928 |
| Resident 12 | 0.8762 | 0.9485 | 0.9109 | 0.9037 |
CONCLUSIONS -
Deep learning presents a promising opportunity for plastic surgery researchers beyond traditional statistics and classical machine learning methods. In our study, we built a deep convolutional neural network capable of discerning between augmented and natural breasts on par with a human expert. We foresee deep learning playing a key role in the advancement of many areas of basic plastic surgery research including: predicting wound healing outcomes, optimizing operative techniques for symmetry and aesthetics, early skin cancer detection, and free flap monitoring.
Back to 2018 Posters