Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics

Yuming Jiang, Kangneng Zhou, Zepang Sun, Hongyu Wang, Jingjing Xie, Taojun Zhang, Shengtian Sang, Md Tauhidul Islam, Jen‐Yeu Wang, Chuanli Chen, Qingyu Yuan, Sujuan Xi, Tuanjie Li, Yikai Xu, Wenjun Xiong, Wei Wang, Guoxin Li, Ruijiang LiCell 2023 · DOI

TL;DR

This paper introduces a novel deep learning radiomics model (HR-Rad-Net) that can non-invasively assess the tumor microenvironment (TME) in gastric cancer from CT scans. This imaging signature not only predicts patient prognosis but also identifies who will benefit from adjuvant chemotherapy and predicts response to immunotherapy, paving the way for personalized cancer treatment.

What Problem Does This Paper Address?

Accurately understanding the tumor microenvironment (TME) is crucial for predicting cancer progression and treatment response. The TME, a complex ecosystem of cells, blood vessels, and signaling molecules surrounding a tumor, significantly influences how a cancer behaves and whether it will respond to therapies like chemotherapy or immunotherapy. Traditionally, TME evaluation relies on invasive methods like immunohistochemistry (IHC) on tissue biopsies. However, these methods have limitations: they are invasive, can be prone to sampling bias due to tumor heterogeneity, and may not always be feasible, especially in neoadjuvant settings or for metastatic disease. This creates a need for non-invasive methods to assess the TME, enabling longitudinal monitoring and more precise treatment decisions.

Methods

The core of this research is the development of a deep learning radiomics model named "HR-Rad-Net." This model ingeniously combines two powerful approaches:

  • Deep Learning (HR-Net): A convolutional neural network architecture (HR-Net) is used to automatically learn complex feature representations directly from CT images. To enhance its ability to capture multi-scale image features, a "squeeze and excitation" (SE) module was incorporated.
  • Radiomics: Hand-crafted radiomics features are extracted from the CT images. These features quantify various aspects of the tumor's texture, shape, and intensity.

These two components are then fused in a fully connected layer to generate a final prediction, creating an "imaging-based TME classifier" or "deep learning radiomics signature" (DLRS). The model was trained on a large dataset of gastric cancer patients and validated across multiple independent cohorts. The researchers also employed techniques like data augmentation and guided Grad-CAM for model interpretation, visualizing which image regions were most important for prediction.

Datasets

The study utilized data from a total of 2,686 patients across six independent cohorts:

  • Training Cohort: 398 patients from Nanfang Hospital, Southern Medical University (SMU).
  • Internal Validation Cohorts: 196 patients (SMU) and 602 patients (SMU).
  • External Validation Cohorts: 101 patients (Sun Yat-sen University Cancer Center - SYSUCC) and 1,068 patients (SYSUCC).
  • Immunotherapy Cohort: 321 patients with advanced gastric cancer treated with anti-PD-1 immunotherapy at Nanfang Hospital and Guangdong Provincial Hospital of Chinese Medicine.

Additionally, radiogenomics analysis was performed using the TCGA/TCIA-STAD dataset, which includes genomic/transcriptomic data and matched CT images for 42 gastric cancer patients.

Key Insights and Ideas

  • Bridging the Gap: The study successfully bridges the gap between imaging and the complex biological reality of the TME, offering a non-invasive window into a critical aspect of cancer biology.
  • Synergy of DL and Radiomics: Combining deep learning's ability to learn features with radiomics' quantitative descriptors proved more powerful than either approach alone for TME classification.
  • Beyond Prognosis: The DLRS is not just a prognostic marker; it actively predicts response to specific therapies, highlighting its clinical utility in treatment selection.
  • Personalized Chemotherapy: The DLRS can identify patients who will not benefit from adjuvant chemotherapy, potentially sparing them unnecessary toxicity and guiding more targeted treatment strategies.
  • Immunotherapy Enhancement: While PD-L1 expression (CPS) is a known biomarker for immunotherapy response, the DLRS offers complementary information and, when combined, significantly improves predictive accuracy.

Key Results

  • Accurate TME Classification: The HR-Rad-Net model achieved high accuracy in predicting TME status across training and validation cohorts, with AUCs ranging from 0.909 to 0.937.
  • Prognostic Value: The DLRS was an independent prognostic factor for both disease-free survival (DFS) and overall survival (OS) in gastric cancer patients, outperforming traditional clinicopathological variables.
  • Chemotherapy Benefit Prediction: Patients with high DLRS scores in stage II/III gastric cancer significantly benefited from adjuvant chemotherapy, whereas low DLRS patients did not. For low DLRS patients in stage II, chemotherapy was even associated with worse DFS.
  • Immunotherapy Response Prediction: In patients treated with anti-PD-1 immunotherapy, high DLRS was associated with a substantially higher objective response rate (57.8% vs. 14.2%) and improved progression-free survival. The DLRS also improved predictive accuracy over PD-L1 CPS alone and, when combined, yielded an AUC of 0.783 for predicting response.

Limitations and Future Work

The authors acknowledge several limitations:

  • Retrospective Design: The study is retrospective, which inherently carries the risk of selection bias.
  • Clinical Trial Validation: The predictive effect of the imaging model needs validation in prospective, randomized clinical trials.
  • Geographic Specificity: Patients were primarily from one geographic region in China, suggesting that the distribution of clinical and pathological characteristics might differ in other populations.

Future work will focus on:

  • Prospective Validation: Confirming the generalizability and reproducibility of the imaging signature in larger, diverse populations and across different scanners.
  • Adjunct Tool: Positioning the imaging model as an adjunct tool to complement existing histopathology approaches, especially when tissue is unavailable or insufficient.
  • Extension to Other Cancers: Exploring the applicability of this approach to other solid tumor types.

Summary updated

Abstract

The tumor microenvironment (TME) plays a critical role in disease progression and is a key determinant of therapeutic response in cancer patients. Here, we propose a noninvasive approach to predict the TME status from radiological images by combining radiomics and deep learning analyses. Using multi-institution cohorts of 2,686 patients with gastric cancer, we show that the radiological model accurately predicted the TME status and is an independent prognostic factor beyond clinicopathologic variables. The model further predicts the benefit from adjuvant chemotherapy for patients with localized disease. In patients treated with checkpoint blockade immunotherapy, the model predicts clinical response and further improves predictive accuracy when combined with existing biomarkers. Our approach enables noninvasive assessment of the TME, which opens the door for longitudinal monitoring and tracking response to cancer therapy. Given the routine use of radiologic imaging in oncology, our approach can be extended to many other solid tumor types.
BibTeX
@inproceedings{jiang2023non,
  author = {Yuming Jiang and Kangneng Zhou and Zepang Sun and Hongyu Wang and Jingjing Xie and Taojun Zhang and Shengtian Sang and Md Tauhidul Islam and Jen‐Yeu Wang and Chuanli Chen and Qingyu Yuan and Sujuan Xi and Tuanjie Li and Yikai Xu and Wenjun Xiong and Wei Wang and Guoxin Li and Ruijiang Li},
  title = {{Non-invasive tumor microenvironment evaluation and treatment response prediction in gastric cancer using deep learning radiomics}},
  booktitle = {Cell},
  year = {2023},
  doi = {10.1016/j.xcrm.2023.101146},
}
Figure