Research Report on Immunotherapy Efficacy Prediction for Non-Small Cell Lung Cancer: A Study of Multi-View Adaptive Weighted Graph Convolutional Networks

Background Introduction

Lung cancer is a highly prevalent and poorly prognostic malignant tumor with a persistently high mortality rate. Among all lung cancer patients, Non-Small Cell Lung Cancer (NSCLC) accounts for approximately 85%. As an emerging treatment approach, cancer immunotherapy offers new treatment options for cancer patients. However, immunotherapy is expensive, and only about 20% to 50% of patients can achieve satisfactory results. Additionally, side effects such as immunogenic pneumonia and hepatitis may occur during treatment. Therefore, predicting the efficacy of immunotherapy before patients undergo treatment is of great significance.

In recent years, radiomics based on machine learning has shown potential in predicting the efficacy of NSCLC immunotherapy. Radiomics features have been proven to be effective surrogate markers for predicting immunotherapy efficacy. However, most studies only consider the radiomics features of individual patients, ignoring the interrelationships between patients. Furthermore, they typically concatenate different features as a single-view model input, failing to fully account for the complex correlations among various types of features.

Source of the Paper

The paper titled “Immunotherapy Efficacy Prediction for Non-Small Cell Lung Cancer Using Multi-View Adaptive Weighted Graph Convolutional Networks” was co-authored by Qiong Wu, Jun Wang (IEEE Member), Zongqiong Sun, Lei Xiao, Wenhao Ying, and Jun Shi (IEEE Member). It was published in the IEEE Journal of Biomedical and Health Informatics, Volume 27, Issue 11, in November 2023.

Research Process

Data Source and Preprocessing

This study’s dataset includes data retrospectively collected from 107 NSCLC patients between January 2018 and October 2020. These patients underwent CT imaging scans three days before receiving anti-PD-1 immunotherapy. The evaluation of immunotherapy efficacy was based on the Immune-Related Response Evaluation Criteria In Solid Tumors (iRECIST) standards, assessed by experienced radiologists, and received corresponding ethical approval.

Radiomics Feature Extraction

Using the guidelines from the Image Biomarker Standardization Initiative (IBSI) and employing the open-source software package Pyradiomics, multiple categories of radiomics features were extracted from the regions of interest (ROIs) in CT images. These images were processed using various filters such as wavelet filter, square, square root, logarithm, exponential, gradient, and local binary pattern 2D (LBP2D) filters, and the features were grouped by image preprocessing filter type.

Multi-View Graph Construction

In a multi-view context, we use ( X_m ) to represent the feature matrix in the m-th view and ( A_m ) to represent the adjacency matrix in the m-th view. Based on baseline patient features (e.g., age, pathological type, and PD-L1 expression), we calculated the similarity between patients and constructed multi-view graphs. The specific similarity calculation method and formula are as follows:

A_m(i, j) = s(x_{i, m}, x_{j, m}) \sum_{k=1}^{k} J(p_i^{(k)}, p_j^{(k)}),
s(x_i, x_j) = exp(-\frac{c^2(x_i, x_j)}{2\sigma^2}),
c(x_1, x_2) = 1 - \frac{cov(x_1, x_2)}{\sigma(x_1) \sigma(x_2)},
J(p_i^{(k)}, p_j^{(k)}) = \left\{
    \begin{array}{ll}
        1 & \text{if } |p_i^{(k)} - p_j^{(k)}| < \tau \\
        0 & \text{otherwise}
    \end{array}
\right.

Proposed Method: Multi-View Adaptive Weighted Graph Convolutional Network (MVAW-GCN)

Network Structure

We propose a Multi-View Adaptive Weighted Graph Convolutional Network (MVAW-GCN) consisting of three channels, each containing two graph convolution layers and one separable graph convolution operation layer. The output includes both view-shared and view-specific embeddings, which are subsequently fused using an attention mechanism.

The graph convolution formula is:

H_m^{(l)} = ReLU(D_m^{(∼)} ^{-1/2} A_m^{(∼)} D_m^{(∼)} ^{-1/2} H_m^{(l-1)} W_m^{(l)})

The formula for the separable graph convolution operation is:

H_m = ReLU(D_m^{(∼)} ^{-1/2} A_m^{(∼)} D_m^{(∼)} ^{-1/2} H_m^{(l-1)} W_m^s) + ReLU(D_m^{(∼)} ^{-1/2} A_m^{(∼)} D_m^{(∼)} ^{-1/2} H_m^{(l-1)} W_c)

View Fusion Module

The view fusion module assigns adaptive weights to each view using an attention mechanism. The specific attention value calculation formula is:

\tilde{h}_{i,m} = tanh(W_{att} h_{i,m} + b_{att}),
\alpha_{i,m} = softmax(v^t \tilde{h}_{i,m})

The view weight is:

h = \sum_m^M \lambda_m H_m

Loss Function

Considering the view-shared and view-specific information among multiple views, we introduce two types of loss functions: consistency loss (L_c) and diversity loss (L_d).

Consistency loss:

L_c = \frac{1}{M} \sum_{m=1}^{M} ||H_m^c - H^c||_F^2,
H^c = \frac{1}{M} \sum_{m=1}^{M} H_m^c

Diversity loss:

L_d = \sum_{m_1=1}^{M} \sum_{m_2=1}^{m_2≠m_1} HSIC(H_{m1}^s, H_{m2}^s)

The final loss function:

L = L_p + \beta_1 L_c + \beta_2 L_d

Results

We validated the proposed method on a dataset comprising 107 NSCLC patients, including 52 effective treatment patients and 55 ineffective treatment patients. Our method achieved an accuracy of 77.27% and an AUC of 0.7780 in efficacy prediction, demonstrating its effectiveness in predicting NSCLC immunotherapy efficacy.

1. Accuracy Comparison Experiments: Our method, using multiple views of radiomics features, outperformed single-view and simple feature concatenation methods in terms of accuracy, sensitivity, specificity, and F1 score, demonstrating the advantages of multi-view learning.
2. Effectiveness Analysis of Loss Functions: By simultaneously considering consistency and diversity losses, our method achieved the best accuracy and AUC, indicating the importance of considering both shared and specific information among views.
3. Effectiveness of the Attention Mechanism: Experiments showed that during training, the attention mechanism could dynamically adjust the weights for each view, leading to more accurate classification results.

Significance of the Research

This study proposed a multi-view adaptive weighted graph convolutional network method, which effectively improved the accuracy of NSCLC immunotherapy efficacy prediction by comprehensively considering imaging and non-imaging information. This research not only provides new ideas for the application of machine learning and radiomics in the medical field but also offers important references for clinical medical decision-making, helping to evaluate whether patients can benefit from immunotherapy before treatment, thereby reducing unnecessary treatment costs and the risk of side effects.

Future Prospects

Despite the significant achievements in predicting NSCLC immunotherapy efficacy, there are still some limitations, such as sample size limitation and single-center data issues. Future research could consider collecting and validating multi-center data, as well as further exploring more comprehensive biomarkers, such as the impact of PD-L1 expression.

This study provides strong theoretical and experimental support for predicting NSCLC immunotherapy efficacy using multi-view graph convolutional networks and lays the foundation for further research in this field.