Automatic discrimination of neoplastic epithelium and stromal response in breast carcinoma

Abstract

Background and objectives: In breast carcinoma, epithelial-stromal interactions play a pivotal role in tumor formation and progression, and it must be accurately assessed for a correct extraction of predictive and prognostic biomarkers. Evaluation of preoperative (baseline) neoplasia/stroma ratio and the enumeration of tumor infiltrating lymphocytes (TIL) represent only two conditions in which precise discrimination of cancer epithelium and stromal reaction are relevant. However, subjectivity and expertise of the operators may lead to different degrees of assessment.

Methods: In this paper, we present a fully automated method for the discrimination between neoplastic epithelium and stromal reaction in breast carcinoma. Starting from cell nuclei, the proposed method implements computer vision strategies to split the neoplastic epithelium tissue from the stromal reaction.

Results: The algorithm is tested on 100 H&E (hematoxylin and eosin) stained images representative of 10 different cases of invasive carcinoma. The algorithm performance in the detection of neoplastic epithelium (compared to manual annotations by an expert pathologist) gave a F1_SCORE of 0.8894 and mean jaccard_INDEX of 0.8481.

Conclusions: To the best of our knowledge, the proposed method is the first fully automated algorithm for the discrimination between neoplastic epithelium and stromal reaction in H&E stained images of breast tissue. The proposed approach paves the way for an automated and quantitative analysis of predictive and prognostic biomarkers in breast carcinoma.

Keywords: Artificial neural networks; Automatic segmentation; Breast carcinoma; Computer-aided image analysis; Digital pathology.