# A computer vision method to evaluate tumor-infiltrating lymphocytes and multiparametric modeling of neoadjuvant systemic therapy response in breast cancer

## TL;DR

This study uses computer vision and machine learning to predict breast cancer treatment response by analyzing tumor-infiltrating lymphocytes in pre-treatment biopsies.

## Contribution

A novel computer vision method and multiparametric model for predicting neoadjuvant therapy response in breast cancer using TIL spatial features.

## Key findings

- ML models using clinical and graph-based features achieved high predictive accuracy (AUC up to 0.955).
- Ensemble models integrating clinical and graph features outperformed individual models in predicting pCR.
- Significant differences in performance were observed for triple-negative breast cancer models.

## Abstract

Neoadjuvant systemic therapy (NST) is often used to treat locally advanced breast cancer (BC) or patients with early-stage BC at high risk for micrometastatic spread. Pathological complete response (pCR) to NST in BC is associated with excellent prognostic outcomes; however, rates vary significantly. Tumor-infiltrating lymphocytes (TILs) are associated with NST response, suggesting potential as predictive biomarkers.

To develop a computer vision approach to quantify spatial TIL parameters and a multiparametric machine learning (ML) model for predicting NST response.

Retrospective, single institution study of 411 BC patients, combining clinical and graph-level pre-treatment histopathology data to predict response to NST using ML.

Pre-treatment core needle biopsies were prepared, stained with hematoxylin and eosin, and digitized into whole slide images. Convolutional neural networks were applied to segment and classify regions of invasive carcinoma and TILs. Spatial features were extracted based on the coordinates of the TILs within invasive regions, including metrics from Delaunay triangulation, Voronoi diagram analysis, and minimum spanning trees, as well as features capturing cell density and nuclear count. Clinicopathological features were incorporated to support multiparametric modeling. Multiple ML classification models were trained to predict pCR. Logistic regression, K-nearest neighbor, support vector, random forest, Gaussian Naïve Bayes, and extreme gradient boosting models were tested, and model performances were reported.

ML models using clinical and graph-based features achieved high predictive accuracy. The best performing graph feature model reached an area under the receiver operating characteristic curve (AUC) of 0.924. Ensemble models integrating clinical and graph features showed the highest performance, with an AUC of 0.955. Notably, for triple-negative BC, significant differences in predictive performance were demonstrated between clinical and graph feature models (p = 0.026) and between clinical and ensemble models (p = 0.006).

Multiparametric modeling utilizing clinicopathological and graph features obtained from TILs is associated with pCR in BC patients treated with NST.