Amrik Majumdar

Begun 2025 · with a collaborator

Lung Cancer Research

Repeated screening gives a series of CT scans, three-dimensional X-ray images, taken over time. Can a computer model use changes across those scans to predict lung-cancer stage, how far the cancer has spread, more accurately than a model using one scan?

The study compared predictions from three scans with predictions from one scan. It used 236 National Lung Screening Trial participants whose cancer was confirmed by biopsy, a tissue test. Each had three low-dose CT scans taken before treatment.

Both lungs from a chest CT in three dimensions, the blood vessels branching inside them, and one dark, round nodule.
A real chest CT: both lungs, their blood vessels and a nodule a radiologist outlined (darkest). Drag it, or use the arrow keys, to turn it yourself. A public scan, not from the study: LIDC-IDRI, The Cancer Imaging Archive, CC BY 3.0.

Three scans versus one

  1. One scan

    A single CT scan shows a tumor at one moment. The comparison model, called the baseline, predicted the stage using only that scan.

  2. Three scans in order

    The three-scan model reads a participant’s scans in order, all taken before treatment. It can use the changes between scans to predict the stage.

  3. Time between scans

    The model also uses the time between scans, so it does not treat unequal gaps as equal. Across the study group, the median gap was 12 months: half the gaps were shorter or equal, and half were longer or equal.

  4. The model

    The model reads the three scans together and predicts the stage. It uses a 3D residual network, a neural network that reads three-dimensional images and carries earlier information through later layers. Dual attention uses two mechanisms to focus on relevant features. Interval-aware temporal encoding tells the network the order of the scans and the time between them.

Results

One scan, 0.82. Three scans in order, 0.86. The abstract’s overall-stage macro-AUC averages how well the model ranks cases across the stage classes. The figure runs from 0.5 (chance) to 1.0 (perfect ranking).

The three-scan model scored 0.86 versus 0.82 for the single-scan model, a small but statistically significant improvement (p < 0.05). Macro-AUC is a ranking score, not the percentage of correct predictions.

The largest gains were in N stage, which describes lymph-node involvement (+0.13), and T stage, which describes the tumor itself (+0.11). Attention maps show which areas the model focused on. They highlighted the tumor’s edge and the tissue just outside it. The abstract describes this as consistent with cancer invading nearby tissue.

For this study group, changes across scans helped predict stage. This is a meeting abstract, not a model validated for use in patient care. Testing on data from outside the study was still underway when the abstract was published.

Meeting abstract 8018

Lung cancer stage prediction using longitudinal CT imaging compared with single-timepoint assessment.

Amrik Majumdar and a collaborator

Journal of Clinical Oncology 44, 16_suppl · published online May 27, 2026 · doi:10.1200/JCO.2026.44.16_suppl.8018

Method

Evaluation
To test the model, the data was split into five groups, with each group reserved for testing in turn. This is nested five-fold cross-validation: settings are chosen using separate splits within each round’s training data. The test group is used only to score the model, not to choose its settings.
Stage labels
The AJCC seventh-edition staging system supplied the stage labels. These were pathological stages, assessed from tissue, when a surgical specimen was available. Otherwise, they were clinical stages, assessed without a surgical specimen.
Read the abstract and method details