Use of AI-based second harmonic generation digital pathology for predicting patient survival for triple negative breast cancer patients.

Abstract

e12526 Background: Extracellular matrix (ECM) in the stromal region has been known to be associated with tumorigenesis and metastasis in breast cancers. About 10-20% of all breast cancers are triple-negative breast cancers (TNBC), considered to be more aggressive with poorer prognosis than other types of breast cancer. AI-based Second Harmonic Generation (SHG) digital pathology, which is a fully quantitative fibrosis assessment, has been widely used innon-alcoholic steatohepatitis (NASH) and is FDA approved primary endpoint in NASH phase 2 clinical trials. In oncology, AI-based SHG digital pathology has demonstrated its relevance in survival prognosis for hepatocellular carcinoma and renal cell carcinoma after surgical treatment, based on analysis of tumor and non-tumor tissue collagen parameters. In this study, we study its potential applications for survival prognosis of TNBC patients. Methods: 68 patients with TNBC were treated by breast-cancer excision surgery. Tumor and non-tumor tissue from excised mass was imaged by SHG microscope (Genesis200, HistoIndex Pte. Ltd., Singapore). Regular follow-up was performed for patients post-operatively. A total of 33 collagen morphology features for collagen strings were quantified, such as length/width of strings, number of long/short/thick/thin strings, from tumour and non-tumour tissues. We used these collagen parameters to build two survival prediction models (RFS-index and OS-index) to predict patient’s recurrence-free survival (RFS) and overall survival (OS) years. The models were validated using the leave-one-out method. Results: Both RFS-index and OS-index were created using 10 specific collagen parameters chosen by sequential selection methods from overall 33 collagen parameters evaluated. 9 out of 10 parameters were selected from non-tumour tissue. The RFS-index can differentiate patients with RFS≥3 years (n = 36) and RFS < 3 years (p < 0.001) with cut-off value of RFS-index = 0.50. The OS-index can differentiate patients with OS≥5 years (n = 42) and OS < 5 years (p < 0.001) with cut-off value of RFS-index = 0.55. The log-rank test showed RFS-index (p = 0.026) and OS-index (p < 0.001) can be used for prediction of disease-free and overall survival based on collagen parameters. Conclusions: SHG based AI-aided quantitative assessment of ECM collagen has been shown to correlate with survival rates of TNBC patients. This is a proof of concept study applying AI based digital pathology in helping predict survival rates for TNBC patients based on tumor and non-tumor tissue ECM collagen parameters. This could help identify patients with a higher risk of recurrence of disease and lower overall survival rates. Such patients can be followed-up more carefully, and be considered for earlier treatment interventions to improve their survival outcomes.