AI Radiomics: Predicting Lung Cancer Immunotherapy Response

This article provides a comprehensive overview of how radiomics and artificial intelligence (AI) analyze standard Computed Tomography (CT) scans to predict whether a lung tumor will respond to PD-1 checkpoint immunotherapy. By extracting and evaluating sub-visual texture features that the human eye cannot perceive, AI models can decode the physical structure of the tumor microenvironment. Below, we explore the step-by-step process of image segmentation, feature extraction, and machine learning classification that enables oncologists to personalize lung cancer treatment without relying solely on invasive tissue biopsies.

The Challenge of PD-1 Immunotherapy

PD-1 checkpoint inhibitors have revolutionized the treatment of non-small cell lung cancer (NSCLC) by helping the body's immune system recognize and attack cancer cells. However, only about 20% to 40% of patients respond to this therapy. Currently, doctors rely on invasive tissue biopsies to measure PD-L1 protein levels to predict treatment success. Because lung tumors are heterogeneous—meaning different parts of the same tumor have different biological characteristics—a single needle biopsy often misses the complete picture. Radiomics solves this by analyzing the entire three-dimensional tumor using routine CT scans.

Step 1: Image Acquisition and Tumor Segmentation

The process begins with a standard diagnostic CT scan. Before the AI can analyze the tumor, the exact boundaries of the lesion must be defined in a process called segmentation.

  • Region of Interest (ROI): Automated deep learning algorithms or radiologists trace the 3D volume of the tumor, known as the ROI.
  • Peritumoral Zone: The segmentation often includes a margin of healthy-looking tissue immediately surrounding the tumor. This peritumoral zone is critical because it is where immune cells first attempt to infiltrate the cancer.

Step 2: Extracting Sub-Visual Texture Features

To the human eye, a CT scan is a collection of greyscale shadows. To a computer, a CT scan is a massive 3D grid of numerical values called voxels (3D pixels), where each number represents tissue density (measured in Hounsfield Units). Radiomics software scans the segmented tumor and uses mathematical formulas to extract thousands of "sub-visual" features. These are broken down into three main categories:

  • First-Order Features: These analyze the basic distribution of voxel intensities without looking at their spatial location. They measure metrics like average density, skewness, and uniformity.
  • Second-Order Features (Texture): These measure the spatial relationship between adjacent voxels. Using tools like the Grey-Level Co-occurrence Matrix (GLCM), the software calculates how often pixels of certain brightness levels appear next to each other. This quantifies tumor "roughness," contrast, or smoothness.
  • Higher-Order Features: Algorithms apply mathematical filters (like wavelet transforms) to the image to highlight repeating patterns, edges, and fractal dimensions at different scales.

Step 3: Linking CT Textures to Tumor Biology

The reason these sub-visual textures matter is that they directly reflect the underlying biology of the lung tumor. PD-1 immunotherapy requires active immune cells (T-cells) to be present inside the tumor to work.

  • Heterogeneous Textures: A tumor that shows chaotic, irregular voxel patterns on a CT scan often correlates with regions of rapid cell death (necrosis), abnormal blood vessel growth, and pockets of immune cell infiltration.
  • Homogeneous Textures: A uniform, smooth texture on the scan may indicate a dense, solid tumor with very few infiltrating immune cells (an "immune-desert" tumor), which is unlikely to respond to PD-1 inhibitors.

Step 4: AI Classification and Prediction

Extracting thousands of texture features creates too much data for a human to interpret. This is where artificial intelligence and machine learning come in.

First, feature selection algorithms filter out redundant data and noisy variables caused by variations in different CT scanners. This leaves only the most stable, predictive features. Next, machine learning models—such as Random Forests, Support Vector Machines, or Deep Neural Networks—are trained on historical patient data. The AI looks at the radiomic signature of thousands of past patients and learns which specific combinations of sub-visual textures belonged to patients whose tumors shrank after PD-1 therapy, and which belonged to patients whose cancer progressed.

Once trained, the AI model can take the CT scan of a newly diagnosed lung cancer patient, extract the sub-visual texture features, and output a "radiomic score." This score gives the oncologist a clear probability of whether the tumor will respond to PD-1 checkpoint immunotherapy, allowing doctors to fast-track effective treatments and spare non-responders from severe side effects and delays in alternative care.