Abstracts

RETURN TO ABSTRACT LISTING


2160. Weakly Supervised Deep Learning Approach to Breast MRI Assessment
Authors * Denotes Presenting Author
  1. Shawn Sun *; Columbia University Irving Medical Center/New York Presbyterian
  2. Michael Liu; Columbia University Irving Medical Center/New York Presbyterian
  3. Cara Swintelski; Columbia University Irving Medical Center/New York Presbyterian
  4. Maham Siddique; Columbia University Irving Medical Center/New York Presbyterian
  5. Sachin Jambawalikar; Columbia University Irving Medical Center/New York Presbyterian
  6. Richard Ha; Columbia University Irving Medical Center/New York Presbyterian
Objective:
Deep learning techniques have been to shown to have high performance and potential for the analysis of breast MRI. A weakly supervised approach to applying deep learning techniques does not require pixel by pixel segmentation of images. The purpose of this study is to test the feasibility of a weakly supervised approach in assessing breast MRI images without pixel level segmentation and to evaluate whether this approach can improve the specificity of breast MRI lesion classification.

Materials and Methods:
An IRB approved study was performed. We used both an internal and external dataset. The internal dataset consisted of 302 patients from our institution and the publically available external dataset consisted of 136 patients totaling 278,685 image slices and 46,707 image slices containing either tumor or metastatic lymph node confirmed by tissue pathology. The weakly supervised network was based on the Resnet-101 architecture (1-2) and network weights were initialized randomly. Training was implemented using the Adam optimizer and a final SoftMax score threshold of 0.5 was used for two class classification (malignant or benign). The total dataset of 278,685 gray scale image, 79,871 (85%) images were used for training and validation while 13,024 (15%) images were separated as a hold out testing dataset. Of the 13,024 images in the hold out testing dataset contained 11,498 (88%) benign images, 1,003 (8%) tumor images and 523 (4%) metastatic lymph node images. Model performance was assessed in terms of accuracy, sensitivity, specificity and area under the receiver operating curve (AUC). In addition, class activation mapping was created to assess the areas in an image that were most significant in a network’s detection of abnormal tissue.

Results:
The weakly supervised network achieved an AUC of 0.92 (SD± 0.03) in distinguishing malignant from benign images in breast MRI images in the hold out testing dataset. The model had an accuracy of 94.2% (SD±3.4) with a sensitivity and specificity of 74.4% (SD± 8.5) and 95.3% (SD± 3.3) respectively.

Conclusion:
It is feasible to use a weakly supervised based CNN approach to assess breast MRI images without the need for pixel by pixel segmentation yielding a high degree of specificity in lesion classification. This method may enable generation of large datasets without the need for manual annotation and has the potential for clinical application in breast MRI interpretation.