Contrastive Representation Learning for Efficient Lung Cancer Survival Prediction from CT Imaging

Principal Investigator

Name
Aldo Pacchiano Camacho

Degrees
PhD

Institution
Trustees of Boston University

Position Title
Assistant Professor

Email
pacchian@bu.edu

About this CDAS Project

Study
NLST (Learn more about this study)

Project ID
NLST-1532

Initial CDAS Request Approval
Sep 28, 2026

Title
Contrastive Representation Learning for Efficient Lung Cancer Survival Prediction from CT Imaging

Summary
Lung cancer remains the most frequently diagnosed cancer globally. Recent studies have
demonstrated the potential of artificial intelligence models to estimate patient survival probability [1, 2]. However, a significant challenge persists in processing high-dimensional imaging data such as CT scans while developing computationally efficient machine learning pipelines for Cox proportional hazard modeling [3].

Over-parameterized models like convolutional neural networks (CNNs) present particular computational obstacles when employing second-order optimization methods, which are standard for fitting Cox proportional hazard models. To address this limitation, contrastive learning approaches show promise [4] by generating lower-dimensional embeddings that capture the most informative aspects of CT imaging data. This enables the application of simpler, more computationally tractable models using second-order methods such as Newton-Raphson algorithms.

The NLST dataset is essential for achieving these objectives, as it provides a comprehensive public resource containing both imaging and survival data. We respectfully request access to this dataset to train and rigorously evaluate our proposed models.

Bibliography
[1] Torres, Felipe Soares, et al. "End-to-end non–small-cell lung cancer prognostication using deep learning applied to pretreatment computed tomography." JCO Clinical Cancer Informatics 5 (2021): 1141-1150.
[2] Sako, Chiharu, et al. "Deep-Learning Serial CT Prediction of Survival in Immunotherapy-Treated Non–Small Cell Lung Cancer. " JAMA Network Open 9.1 (2026): e2555759.
[3] Tarkhan, Aliasghar, et al. "Survival prediction using deep learning." Survival Prediction-Algorithms,Challenges and Applications. PMLR, 2021.
[4] Lee, Dongjoon, Hyeryn Park, and Changhee Lee. "Toward a Well-Calibrated discrimination via survival Outcome-Aware contrastive learning." Advances in Neural Information Processing Systems 37 (2024): 30985-31014.

Aims

Aim 1: Construct an analysis-ready cohort linking low-dose CT imaging with survival outcomes. Using the National Lung Screening Trial (NLST) dataset, we will develop a reproducible pipeline for selecting eligible participants, preprocessing CT scans, and linking imaging data with available demographic, clinical, and survival information. Particular attention will be given to censoring, missing data, cohort definition, and the prevention of information leakage between training, validation, and test sets.

Aim 2: Learn compact and informative representations of lung CT scans through contrastive learning. We will develop and evaluate contrastive learning methods that transform high-dimensional CT scans into lower-dimensional embeddings. The objective is to retain imaging features relevant to patient prognosis while substantially reducing the computational and statistical complexity of downstream survival modeling. We will investigate alternative contrastive objectives, data-augmentation strategies, and embedding dimensions.

Aim 3: Develop and rigorously evaluate computationally efficient survival-prediction models. We will use the learned CT representations as inputs to Cox proportional hazards models that can be fitted efficiently using second-order optimization methods. We will compare the proposed approach with appropriate clinical, imaging, and machine-learning baselines. Performance will be assessed using established survival-analysis measures, including concordance, time-dependent discrimination, and calibration, as appropriate. We will also quantify computational efficiency and conduct ablation and sensitivity analyses to determine which components contribute most strongly to predictive performance.

Aim 4: Disseminate the resulting methods and findings. We intend to publish the results in a peer-reviewed medical-imaging, machine-learning-for-health, or cancer-informatics venue and present the work at an appropriate scientific conference. To support reproducibility, we will release methodological details and analysis code where permitted by the NLST data-use agreement and applicable institutional and ethical requirements. No participant-level NLST data will be redistributed.

Collaborators

Aldo Pacchiano Camacho Trustees of Boston University
Spyridon Konstantinos Dragazis Trustees of Boston University
Skander Jemaa Genentech, Inc