Abstract P1-02-15: Predicting molecular subtypes of breast cancer using

Abstract

Abstract Purpose: Breast cancer is a heterogeneously complex disease. A number of molecular subtypes with distinct biological features lead to different treatment responses and clinical outcomes. Traditionally, breast cancer is classified into subtypes based on gene expression profiles; these subtypes, namely intrinsic subtypes, include luminal A, luminal B, basal-like and HER2-enriched breast cancer. This molecular taxonomy, however, could only be appraised through transcriptome analyses. We aim to classify breast cancer intrinsic subtypes using un-annotated approach using deep learning models.Methods: 388 pathological whole slide images (WSIs) from TCGA-BRCA dataset were downloaded from TCGA-GDC portal. These WSIs underwent patches generation and normalization using PyHIST tool and Macenko algorithm, respectively. Laplacian algorithm was applied to remove patches with blurry areas and pixelated. The remaining patches (n = 1,833,889) were divided into 3 parts for training (70%), testing (5%) and validation (25%). We applied a 2-step transfer learning with 2 pre-trained models, namely ResNet50, ResNet101, Xception and VGG16, which have been trained on another in-house breast cancer pathological image dataset. Results: these four models shown promising classification results of 4 breast cancer intrinsic subtypes with accuracy ranged from 0.68 (ResNet50 model) to 0.78 (ResNet101 model) in both testing and validation sets. The average AUC score for these models were from 0.88 (ResNet50 model) to 0.94 (ResNet101 model), whereas ResNet101_imgnet with “imagenet” weight archived an accuracy of 0.73 and AUC of 0.92. The overall accuracy of patient-wise prediction even shown a higher average accuracy of 0.914. These models’ prediction visualization was also used to demonstrate that the process of model learning was based on pathological cells’ clusters. Conclusion: Our study demonstrated the feasibility and capability of the deep learning model in classifying breast cancer intrinsic subtypes without region of interest annotation, which wound facilitate the clinical applicability of proposed models. Citation Format: Chi-Cheng Huang, Nam Nhut Phan, Ling-Ming Tseng. Predicting molecular subtypes of breast cancer using [abstract]. In: Proceedings of the 2021 San Antonio Breast Cancer Symposium; 2021 Dec 7-10; San Antonio, TX. Philadelphia (PA): AACR; Cancer Res 2022;82(4 Suppl):Abstract nr P1-02-15.