Predicting Relapse in Patients With Triple Negative Breast Cancer (TNBC) Using a Deep-Learning Approach

Abstract

The abundance and/or location of tumor infiltrating lymphocytes (TILs), especially CD8⁺ T cells, in solid tumors can serve as a prognostic indicator in various types of cancer. However, it is often difficult to select an appropriate threshold value in order to stratify patients into well-defined risk groups. It is also important to select appropriate tumor regions to quantify the abundance of TILs. On the other hand, machine-learning approaches can stratify patients in an unbiased and automatic fashion. Based on immunofluorescence (IF) images of CD8⁺ T lymphocytes and cancer cells, we develop a machine-learning approach which can predict the risk of relapse for patients with Triple Negative Breast Cancer (TNBC). Tumor-section images from 9 patients with poor outcome and 15 patients with good outcome were used as a training set. Tumor-section images of 29 patients in an independent cohort were used to test the predictive power of our algorithm. In the test cohort, 6 (out of 29) patients who belong to the poor-outcome group were all correctly identified by our algorithm; for the 23 (out of 29) patients who belong to the good-outcome group, 17 were correctly predicted with some evidence that improvement is possible if other measures, such as the grade of tumors, are factored in. Our approach does not involve arbitrarily defined metrics and can be applied to other types of cancer in which the abundance/location of CD8⁺ T lymphocytes/other types of cells is an indicator of prognosis.

Keywords: immunofluorescence images; machine-learning; relapse prediction; triple negative breast cancer (TNBC); tumor-infiltrating T cells.

FIGURE 1

The flowchart. Overall description of our machine-learning approach.

FIGURE 2

Training images and accuracy. (A,B) Representative patches (64 × 64 pixels) from patients with the poor and good outcome, respectively. While and red pixels represent PanCK-positive (cancer cells) and CD8-positive (CD8⁺ T cells) areas, respectively. (C) Evolution of the accuracy on training patches as a function of Epochs. (D) The fraction of correctly-predicted patients as a function of the cut-off percentage (R_c) of patches that are classified as arising from a patient with the good outcome.

FIGURE 3

Statistics of CD8⁺ pixels in images. (A) The density of CD8⁺ pixels inside cancer-cell islands, i.e., the number of CD8 + pixels divided by the number of PanCK⁺ pixels. (B) CD8-positive and PanCK-positive areas (in μm²) of the two cohorts. The black dash lines in each figure are selected manually to separate the groups (poor- vs. good-outcome) of the patients which give rise to the highest accuracy when comparing to the actual outcome.

FIGURE 4

Examples of patches. Patches that are predicted to be from patients with the poor outcome (upper 50) or good outcome (lower 50). The size of each patch is 64 pixels (64 pixels, with 1 pixel = 10 μm White and red pixels represent PanCK-positive (cancer cells) and CD8-positive (CD8⁺T cells) areas, respectively. For the patches in the upper 50, we still observe many samples have a lot of red pixels (CD8-positive areas), however, compared to patches in the lower 50, most of the red dots are in the tumor stroma (black areas) instead of cancer-cell islands (white areas).