Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LungEvaty: A Scalable, Open-Source Transformer-based Deep Learning Model for Lung Cancer Risk Prediction in LDCT Screening

Johannes Brandt, Maulik Chevli, Rickmer Braren, Georgios Kaissis, Philip Müller, Daniel Rueckert

TL;DR

LungEvaty introduces a scalable, open-source Transformer framework that predicts 1-6 year lung cancer risk from a single LDCT by processing the entire lung volume. It employs a Primus EVA-02–style backbone with a global CLS token and an attention-query token to capture global and local malignancy cues, and it optionally uses Anatomically Informed Attention Guidance (AIAG) to steer attention toward known malignant regions. The model is pre-trained with Masked Image Modeling on NLST data and fine-tuned with a time-discretized survival objective across yearly horizons, incorporating a Cumulative Hazard Layer for isotonic risk outputs. Across standard NLST-based splits, LungEvaty achieves state-of-the-art ROC-AUC and especially PR-AUC with a single modality and without pixel-level supervision, while AIAG provides gains when annotations are available. The framework is data-efficient, open-source, and adaptable for longitudinal or multimodal extensions, though external validation on independent cohorts remains essential for generalization across scanners and populations.

Abstract

Lung cancer risk estimation is gaining increasing importance as more countries introduce population-wide screening programs using low-dose CT (LDCT). As imaging volumes grow, scalable methods that can process entire lung volumes efficiently are essential to tap into the full potential of these large screening datasets. Existing approaches either over-rely on pixel-level annotations, limiting scalability, or analyze the lung in fragments, weakening performance. We present LungEvaty, a fully transformer-based framework for predicting 1-6 year lung cancer risk from a single LDCT scan. The model operates on whole-lung inputs, learning directly from large-scale screening data to capture comprehensive anatomical and pathological cues relevant for malignancy risk. Using only imaging data and no region supervision, LungEvaty matches state-of-the-art performance, refinable by an optional Anatomically Informed Attention Guidance (AIAG) loss that encourages anatomically focused attention. In total, LungEvaty was trained on more than 90,000 CT scans, including over 28,000 for fine-tuning and 6,000 for evaluation. The framework offers a simple, data-efficient, and fully open-source solution that provides an extensible foundation for future research in longitudinal and multimodal lung cancer risk prediction.

LungEvaty: A Scalable, Open-Source Transformer-based Deep Learning Model for Lung Cancer Risk Prediction in LDCT Screening

TL;DR

LungEvaty introduces a scalable, open-source Transformer framework that predicts 1-6 year lung cancer risk from a single LDCT by processing the entire lung volume. It employs a Primus EVA-02–style backbone with a global CLS token and an attention-query token to capture global and local malignancy cues, and it optionally uses Anatomically Informed Attention Guidance (AIAG) to steer attention toward known malignant regions. The model is pre-trained with Masked Image Modeling on NLST data and fine-tuned with a time-discretized survival objective across yearly horizons, incorporating a Cumulative Hazard Layer for isotonic risk outputs. Across standard NLST-based splits, LungEvaty achieves state-of-the-art ROC-AUC and especially PR-AUC with a single modality and without pixel-level supervision, while AIAG provides gains when annotations are available. The framework is data-efficient, open-source, and adaptable for longitudinal or multimodal extensions, though external validation on independent cohorts remains essential for generalization across scanners and populations.

Abstract

Lung cancer risk estimation is gaining increasing importance as more countries introduce population-wide screening programs using low-dose CT (LDCT). As imaging volumes grow, scalable methods that can process entire lung volumes efficiently are essential to tap into the full potential of these large screening datasets. Existing approaches either over-rely on pixel-level annotations, limiting scalability, or analyze the lung in fragments, weakening performance. We present LungEvaty, a fully transformer-based framework for predicting 1-6 year lung cancer risk from a single LDCT scan. The model operates on whole-lung inputs, learning directly from large-scale screening data to capture comprehensive anatomical and pathological cues relevant for malignancy risk. Using only imaging data and no region supervision, LungEvaty matches state-of-the-art performance, refinable by an optional Anatomically Informed Attention Guidance (AIAG) loss that encourages anatomically focused attention. In total, LungEvaty was trained on more than 90,000 CT scans, including over 28,000 for fine-tuning and 6,000 for evaluation. The framework offers a simple, data-efficient, and fully open-source solution that provides an extensible foundation for future research in longitudinal and multimodal lung cancer risk prediction.

Paper Structure

This paper contains 7 sections, 1 figure, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments and Results
  4. Discussion and Conclusion
  5. Acknowledgements
  6. Conflict of Interest
  7. Compliance with Ethical Standards

Figures (1)

  • Figure 1: LungEvaty architecture. The encoder consists of Primus (EVA-02–style) Transformer blocks wald2025primus, pretrained with masked autoencoding on NLST national2011national. Our framework's whole-lung processing provides a rich, holistic context in the latent representation. To capture both global lung characteristics and local nodule information from this input, we use two learnable tokens: a CLS token and a multi-head attention query. This architecture is highly scalable for risk prediction while also supporting flexible region guidance via our novel Anatomically Informed Attention Guidance (AIAG) loss, which can leverage optional pixel-level expert annotations.