Computational features of tumor-infiltrating lymphocyte architecture of residual disease after chemotherapy on H&E images as prognostic of overall and disease-free survival for triple-negative breast cancer.

Abstract

584 Background: Approximately 30% of all breast cancers are characterized as triple-negative (TNBC). TNBC typically occurs in younger women and is associated with a poorer prognosis relative to other breast cancer subtypes. High levels of tumor-infiltrating lymphocytes (TILs) in residual disease after Neoadjuvant chemotherapy (NACT) have previously been shown to be associated with better prognosis in TNBC. In this work, we sought to evaluate the prognostic value of computationally derived measures of TIL spatial architecture in residual TNBC after NACT. Methods: H&E-stained samples from 92 patients (pts) with TNBC (41 died, 45 had disease recurrence) and residual disease after NACT were retrospectively collected from 2 sites: Instituto Nacional de Enfermedades Neoplásicas (S1) and University Hospitals (S2). 45 pts (16 deaths, 23 recurrences) from S1 formed the training set and 47 pts (25 deaths, 22 recurrences) from S2 formed the independent validation cohort. Samples were digitized at 20x. Computerized algorithms automatically identified 2 types of nuclei (TILs and non-TILs) and built clusters for each nuclei type based on cell proximity. The spatial arrangement of these clusters was then quantified using network graph metrics. The top 5 features, determined by least absolute shrinkage and selection operator, were used to train a Cox regression model that assigned a risk of death and recurrence to each patient on the training set. The percentile 33 risk score was used as a threshold for stratifying pts on the validation set as either low or high risk. For comparison, we also employed a model based on TIL density alone. Survival analysis was used to evaluate the performance of both approaches on disease-free survival (DFS) and overall survival (OS). Results: Pts in S2 (n=47) identified as “high risk” by the model based on spatial architecture of residual TILs had a significantly shorter survival time. The median OS for pts at high risk was 25 months vs. 55 months for low-risk pts. The median DFS for pts at high risk was 32 months vs 51 months for low-risk pts. Univariable analysis showed this model was prognostic for both OS (Hazard Ratio (HR) = 2.57, 95% Confidence Interval (CI): 1.07-6.16, p=0.03) and DFS (HR=2.38, CI: 1.01-5.62, p=0.04). In contrast, the model based on TIL density was not prognostic for OS (HR=1.24, CI: 0.33-4.63, p=0.73) nor DFS (HR=1.19, CI: 0.32-4.34, p=0.78). Conclusions: A computerized image analysis model based on measurements of spatial arrangement of residual TILs and surrounding cells was found to be prognostic in TNBC pts who received NACT. This method appears to be more prognostic than TIL density alone. Additional multisite validation and multivariable analysis is needed to further establish the independent prognostic utility of TIL based image biomarkers in the post-NACT TNBC.

Artificial intelligence‐based digital scores of stromal tumour‐infiltrating lymphocytes and tumour‐associated stroma predict disease‐specific survival in triple‐negative breast cancer

“…The deep CNNs have been shown to effectively classify and segment tissue types and different kinds of nuclei in the WSIs with a high degree of accuracy, opening vast opportunities for understanding the sTME from novel perspectives [15,[24][25][26][27]. Several researchers have recently proposed computational models to assess the role of different biomarkers such as the absolute mitotic count [28], tumour-infiltrating lymphocyte architecture [29], and spatial organisation of TILs [30].…”