E1855. Benign or Malignant; the Breast Imaging Binary: Artificial Intelligence in Breast Sonography
  1. Aren Vierra ; Santa Clara Valley Medical Center
  2. Ting-Ruen Wei; Santa Clara University
  3. Ran Pang ; Santa Clara Valley Medical Center
  4. Mahesh Patel ; Santa Clara Valley Medical Center
  5. Yuling Yan; Santa Clara University
  6. Young Kang ; Santa Clara Valley Medical Center
To achieve high accuracy for diagnosis of malignant breast lesions at sonography by use of a convoluted neural network (CNN).

Materials and Methods:
Our project is a continuation of the work done by Chang et al. in an attempt to improve accuracy and generalizability by modifying an existing deep learning algorithm and including a larger training data set. In total, images were collected from approximately 1200 patients, 600 from an open source set and 600 from our institution. We utilized the DenseNet121 model pre-trained on the ImageNet database as the starting model and applied transfer learning to transfer that knowledge to breast ultrasound images from an open-source dataset provided by Al-Dhabyani et al. that includes 437 benign and 210 malignant images. Specifically, we removed the last layer of the DenseNet121 model and replaced it with a dropout layer and an output layer that output the probability of the input ultrasound breast image belonging to the malignant class. All layers before the dropout layer were frozen initially to train just the output layer with the open-source dataset and, at a smaller learning rate, we unfroze the entire model and trained it again. The open-source dataset is split, 90% for training and 10% for validation which is used for early stopping. The model performed well on the open-source validation set and was subsequently applied to learn the representation of 1238 unique images; all of which were collected, biopsied, and histologically diagnosed at our institution. 64% of our data set was utilized for training; another 16% of our dataset is used for validation and early stopping. The remaining 20% was used for testing which consisted of 166 benign and 82 malignant images. Classification of benign vs malignant was based on histologic diagnosis, and grouped in a similar fashion to Fujioka et al.

With 68 true positives, 6 false positives, 160 true negatives, and 14 false negatives, the model achieved accuracy of 91.94%, precision of 91.89%, recall of 82.93%, sensitivity of 82.92%, and specificity of 96.39%. Modification of the existing algorithm and additional training improved the accuracy and specificity, compared to the previous work.

Our CNN model demonstrated high accuracy and specificity in the characterization of breast masses detected at sonography. At the time of sonographic work up which is typically performed as a diagnostic exam and outside of screening, a highly specific test to determine benign vs malignant can help in the decision on whether or not to biopsy. Alternatively, an Artificial Intelligence model can act as a second check after initial review of the sonographic images, similar to computer aided detection (CAD).