Prediction of MET Overexpression in Non-Small Cell Lung Adenocarcinomas from Hematoxylin and Eosin Images

Kshitij Ingale, Sun Hae Hong, Josh S. K. Bell, Abbas A. Rizvi, Amy Welch, Lingdao Sha, Irvin Ho, Kunal Nagpal, Aïcha BenTaieb, Rohan P. Joshi, Martin C. StumpearXiv 2023 · DOI · arXiv

Introduction

In the landscape of precision oncology, identifying specific genetic signatures is essential for matching patients with targeted therapies. MET overexpression has emerged as a critical biomarker in non-small cell lung cancer (NSCLC), particularly for novel treatments such as the antibody-drug conjugate telisotuzumab vedotin. However, traditional identification methods—including Next-Generation Sequencing (NGS) and Immunohistochemistry (IHC)—face significant practical hurdles. These include the scarcity of biopsy tissue, high costs, and lengthy turnaround times.

Researchers at Tempus Labs have developed a deep learning approach to address these challenges by identifying MET overexpression directly from standard Hematoxylin and Eosin (H&E) pathology slides. Because H&E slides are routinely generated for nearly every cancer patient during initial diagnosis, this method offers a potential "visual fingerprint" of molecular status without requiring additional tissue or expensive specialized testing.

Methodology

Weakly Supervised Learning Framework

The study utilized a weakly supervised attention-based Multiple Instance Learning (MIL) framework. This architecture is specifically designed for digital pathology, where "gigapixel" whole-slide images (WSIs) are too large for standard neural networks. The model processes these images by:

  1. Tile Extraction: Dividing the WSI into smaller image tiles.
  2. Feature Extraction: Using a ResNet18 backbone, pre-trained on histopathological tasks, to convert tiles into compact feature vectors.
  3. Attention Mechanism: Assigning "attention scores" to individual tiles, allowing the model to weigh the importance of different regions (e.g., tumor vs. stroma) when making a slide-level prediction.

Smooth Labeling and Normalization

To better capture the continuous nature of gene expression, the researchers moved away from simple binary classification. They employed a "smooth labeling" approach, transforming raw RNA expression values (Transcripts Per Million) using a sigmoid function. This improved model calibration and performance. Additionally, the team applied ICC profile transformations for color normalization, ensuring the model remained robust across different digital scanning hardware.

Experimental Results

Performance and Validation

The model was developed using a massive, de-identified database from Tempus Labs, including a development set of 1,367 H&E slides with matched RNA-seq data. The model's performance was validated across several independent sets:

  • Primary Performance: The model achieved a mean ROC-AUC of 0.74 in cross-validation and 0.70 on an independent holdout set.
  • Correlation: A significant Pearson correlation (r = 0.43) was observed between the model’s predicted probabilities and actual MET RNA expression levels.
  • Temporal Stability: Performance remained consistent on a temporal test set of 4,331 slides collected 6–10 months after the training period, suggesting the model is resistant to drift in laboratory protocols.

Biological Relevance and Specificity

Analysis of the model's attention scores confirmed that it naturally prioritized tumor-containing regions over necrosis or immune cells, indicating it learned biologically relevant morphological features. Furthermore, when tested against other common mutations such as EGFR or KRAS, the model’s performance dropped toward random chance, confirming its specificity to MET biology.

Impact of Tissue Type and Hardware

The study revealed that the model performed significantly better on large surgical resections (AUC 0.79) than on small core needle biopsies (AUC 0.67). This suggests that the broader spatial architecture of the tumor provides critical context for molecular prediction. Additionally, while different scanners (Philips vs. Aperio) shifted the raw probability distributions, the rank-ordering of patients remained stable, demonstrating hardware robustness.

Discussion and Clinical Utility

While an AUC of 0.70 indicates the model is not yet a standalone diagnostic replacement for IHC or NGS, its value lies in its role as a prioritization engine. By flagging patients with a high likelihood of MET overexpression, clinicians can prioritize these individuals for confirmatory molecular testing, ensuring that limited tissue samples are used most effectively.

Future work will focus on improving performance in small biopsy specimens and conducting prospective clinical trials to determine if H&E-based pre-screening can significantly reduce the time-to-treatment for NSCLC patients. By turning a routine diagnostic byproduct into a predictive tool, this approach represents a significant step toward more accessible precision medicine.

Summary updated

BibTeX
@misc{ingale2023prediction,
  author = {Kshitij Ingale and Sun Hae Hong and Josh S. K. Bell and Abbas A. Rizvi and Amy Welch and Lingdao Sha and Irvin Ho and Kunal Nagpal and Aïcha BenTaieb and Rohan P. Joshi and Martin C. Stumpe},
  title = {{Prediction of MET Overexpression in Non-Small Cell Lung Adenocarcinomas from Hematoxylin and Eosin Images}},
  year = {2023},
  eprint = {2310.07682},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.07682},
}
Figure