2351. Performance of a Deep Convolutional Neural Network (DCNN) on Indeterminate Calcifications that Underwent Biopsy
Authors * Denotes Presenting Author
  1. Ujas Parikh *; New York University School of Medicine
  2. Laura Heacock; New York University School of Medicine
  3. Eric Kim; New York University School of Medicine
  4. Yiqiu Shen; New York University School of Medicine
  5. Nan Wu; New York University School of Medicine
  6. Linda Moy; New York University School of Medicine
  7. Krzysztof Geras; New York University School of Medicine
To assess the performance of a DCNN to predict malignant calcifications on screening mammography and to identify morphologic features of detected calcifications.

Materials and Methods:
A globally-aware multiple instance classifier (GMIC) is a neural network employing a fusion model that aggregates global and local features to predict malignancy on a screening mammogram and generates saliency maps to localize findings. In this retrospective study, the GMIC was trained on over 220,000 screening mammograms from 2010-2017. The clinical potential of the GMIC was assessed by comparing the model to 14 readers. Each reader provided probability estimates for 720 screening mammograms (62 malignant lesions, 365 benign). The reference standard was pathology reports. The textual report was extracted to obtain morphology of the calcifications determined by the original interpreting radiologist. We cross-referenced the saliency maps to ensure the model's correct localization of the calcifications and thus correlative probability of malignancy for the calcifications. There were 92 cases of biopsy-proven calcifications: 73 benign and 19 malignant (1 IDC, 1 ILC, 17 DCIS). Mean size of calcifications was 12 mm (range 3-37 mm). Performance of the model and readers were assessed using area under the ROC curve (AUC) analysis.

The model achieved an AUC of 0.82 while the average radiologists achieved an AUC of 0.76. All malignant lesions (19/92) were detected by both the model and average radiologists. The model obviated 17 of 73 (23%) benign biopsies without missing any cancers. Similar to radiologists, our model identified, with a high probability, malignant calcifications that were segmental and/or pleomorphic in morphology. Of the 17 cases where a biopsy could have been obviated, 12 (71%) had a heterogeneous morphology. Among the 56 benign calcifications that GMIC recommended to biopsy, 23 (41%) of the calcifications were amorphous.

The overall performance of our model in prediction of suspicious calcifications on mammography is comparable to that of an ensemble of radiologists. The DCNN can obviate 23% of benign biopsies without missing any malignancies.