Measuring changes in the tumor microenvironment before and during treatment with an AI-guided model helped to inform treatment outcomes in patients with rare cancers who received pembrolizumab. When intratrumoral tumor-infiltrating lymphocyte density increased on treatment, these patients tended to have improved progression-free and overall survival outcomes from immune checkpoint inhibition.

The researchers suggested that deep learning could enhance histopathologic studies of the tumor microenvironment as a method for characterizing response to immunotherapy.

“AI-based pathology has the potential to provide clinicians with useful information on both the tumor and its surrounding microenvironment, helping to guide personalized treatment decisions for patients receiving immunotherapy,” said lead study author Aung Naing, MD, Professor of Investigational Cancer Therapeutics at The University of Texas MD Anderson Cancer Center. Research findings were published in the Journal for ImmunoTherapy of Cancer.

Study Methods

The researchers analyzed 256 baseline and 248 on-treatment biopsies from patients with rare tumors (n = 84; 10 cohorts) in a phase II trial of pembrolizumab. There were 10 cohorts in the study for the different tumor types, including squamous cell carcinoma of the skin, small cell malignancies of non-pulmonary origin, adrenocortical carcinoma, medullary renal cell carcinoma, carcinoma of unknown primary, penile carcinoma, vascular sarcoma, germ cell/testicular tumor, paraganglioma/pheochromocytoma, or other rare histologic subtype.

They looked at intratumoral tumor-infiltrating lymphocyte density and tumor content on hematoxylin and eosin (H&E) stained slides. Changes in intratumoral tumor-infiltrating lymphocyte density and tumor content, analyzed with multiplex immunofluorescence, were correlated with survival outcomes.

Assessment was conducted with Lunit SCOPE IO, a deep learning–based whole-slide image analyzer that was developed using data from more than 26 cancer types for broad applicability.

Model Development and Validation

The AI model was trained on a large data set annotated by board-certified pathologists, which was used for cell detection and tissue segmentation. Cell- and tissue-level predictions were validated against pathologist annotations from 152 lung cancer whole-slide images and 315 image patches.

At the tissue level, the model achieved a sensitivity of 0.95 and a specificity of 0.88 for the cancer area and scores of 0.84 and 0.90 for the cancer stroma, respectively. At the cell level, the model identified tumor cells with a precision of 0.84, a recall of 0.84, and an F1 performance score of 0.84. For identifying lymphocytes, the model achieved precision of 0.80, a recall of 0.51, and an F1 performance score of 0.63.

Key Findings

In patients with higher intratumoral tumor-infiltrating lymphocytes, a baseline rate of ≥ 60 cells/mm² (hazard ratio [HR] = 0.49; 95% confidence interval [CI] = 0.25–0.99; P = .046), higher CD8-positive and CD8-positive PD-1–positive cells, and lower FoxP3-positive CD8-positive PD-1–positive T-cell density were all individually considered to be related to a favorable progression-free survival. However, the same trend was not seen in the overall cohort (HR = 0.62; 95% CI = 0.37–1.06; P = .082).

In paired biopsies, an on-treatment increase in intratumoral tumor-infiltrating lymphocytes was associated with a trend toward prolonged progression-free survival (HR = 0.64; 95% CI = 0.40–1.06; P = .084) and was significantly associated with an improvement in overall survival (HR = 0.55; 95% CI = 0.35–1.01; P = .037). An on-treatment increase in intratumoral tumor-infiltrating lymphocytes was also associated with a reduced spatial distance between CD8-positive immune and tumor cells.

Decreased tumor content during treatment with pembrolizumab was significantly associated with improved progression-free (HR = 0.51; P = .019) and overall survival (HR = 0.54; P = .042).

Both increased intratumor tumor-infiltrating lymphocytes with decreased tumor content was significantly associated with a prolonged progression-free (HR = 0.35; P = .009) and overall survival (HR = 0.36; P = .029). Patients with these signals had a median overall survival of 42 months vs 10 months without.

“While this AI-powered approach needs validation, this is an exciting step forward because it shows that meaningful insights can be extracted from routine pathology samples across a diverse group of rare cancers,” Dr. Naing said.

DISCLOSURES: Merck Sharp & Dohme Corp., a subsidiary of Merck & Co., Inc., provided the study drug and funded the study. This work was supported in part by Lunit, which provided AI analysis; the National Cancer Institute at the National Institutes of Health; UT MD Anderson institutional programs; and the National Center for Advancing Translational Sciences at the National Institutes of Health. No custom code was used in this study. For full disclosures of the study authors, visit jitc.bmj.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.