Abstract 5393: Comparison of deep learning approaches applied to hematoxylin and eosin-stained whole slide images from women with benign breast disease to predict risk of developing invasive breast cancer

• Published in: Cancer research (Chicago, Ill.) Ill.), 2023-04, Vol.83 (7_Supplement), p.5393-5393
• Main Authors: Saha, Monjoy, Abubakar, Mustapha, Rohan, Thomas E., Pfeiffer, Ruth M., Duggan, Máire A., Richert-Boe, Kathryn, Figueroa, Jonine D., Almeida, Jonas S., Gierach, Gretchen L.
• Format: Article
• Language: English
• ISSN: 1538-7445; 1538-7445
• DOI: 10.1158/1538-7445.AM2023-5393

Purpose: To compare deep learning (DL) approaches applied to hematoxylin and eosin (H&E)-stained whole slide images (WSIs) from women with benign breast disease (BBD) to predict risk of developing invasive breast cancer (BC). Method: Two deep convolutional neural networks (CNNs) based on a customized 16-layer CNN (known as VGG-16 by Visual Geometry Group, University of Oxford) and an automated CNN (Google’s AutoML) were trained using H&E-stained WSIs to identify distinct histological features on diagnostic BBD biopsies that characterize BBD patients who were (cases, n=347) and were not (controls, n=347) subsequently diagnosed with invasive BC. The CNNs consisted of multiple convolutions, max pooling, fully connected, etc., layers. To incorporate our data into the VGG network, we customized the network architecture and hyperparameters to enhance the classification performances. For AutoML, we used the system's default network with standard hyperparameters. The trained model was then tested on a held-out set of 140 patients (70 cases and 70 controls). The quantitative performance was evaluated using accuracy (ACC), sensitivity (SE), precision (PR), area under the receiver operating characteristic curve (AUROC), etc. For qualitative results, we generated heatmaps using weights and feature maps from the final convolution layer of our customized CNN. Heatmaps were superimposed onto original H&E images to highlight different unique features (such as pattern, texture, color, and morphology). Results: We found both deep learning methods to demonstrate remarkable ability in predicting case-control status in the held-out set (AUROC= 90% and 89% for customized CNN and AutoML, respectively). However, our customized CNN outperformed AutoML in terms of ACC (83.57% (95% confidence interval (CI): 76-89%) vs 77.86% (95%CI: 70-84%), respectively); SE (82.85% (95%CI: 72-91%) vs 77.86% (95%CI: 70-84%), respectively); PR (84.05% (95%CI: 73-92%) vs 81.97% (95%CI: 70-91%), respectively); F1 score (83.45% (95%CI: 76-89%) vs 76.34% (95%CI: 68-83%), respectively); as well as error rates (0.16% (95%CI: 0.11-0.24%) vs 0.22% (95%CI: 0.16-0.30%), respectively). Heatmaps revealed specific stromal and epithelial features that were distinct between case and control images. Conclusion: By using routinely available H&E-stained WSIs, we developed a customized CNN that outperformed AutoML in distinguishing future BC cases from controls in a BBD population. The qualitative results identified s