Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LinGuinE: Longitudinal Guidance Estimation for Volumetric Tumour Segmentation

Nadine Garibli, Mayank Patwari, Bence Csiba, Yi Wei, Kostantinos Sidiropoulos

TL;DR

LinGuinE (Longitudinal Guidance Estimation), a PyTorch framework that combines image registration and guided segmentation to deliver lesion-level tracking and volumetric masks across all scans in a longitudinal study from a single radiologist prompt, is introduced.

Abstract

Longitudinal volumetric tumour segmentation is critical for radiotherapy planning and response assessment, yet this problem is underexplored and most methods produce single-timepoint semantic masks, lack lesion correspondence, and offer limited radiologist control. We introduce LinGuinE (Longitudinal Guidance Estimation), a PyTorch framework that combines image registration and guided segmentation to deliver lesion-level tracking and volumetric masks across all scans in a longitudinal study from a single radiologist prompt. LinGuinE is temporally direction agnostic, requires no training on longitudinal data, and allows any registration and semi-automatic segmentation algorithm to be repurposed for the task. We evaluate various combinations of registration and segmentation algorithms within the framework. LinGuinE achieves state-of-the-art segmentation and tracking performance across four datasets with a total of 456 longitudinal studies. Tumour segmentation performance shows minimal degradation with increasing temporal separation. We conduct ablation studies to determine the impact of autoregression, pathology specific finetuning, and the use of real radiologist prompts. We release our code and substantial public benchmarking for longitudinal segmentation, facilitating future research.

LinGuinE: Longitudinal Guidance Estimation for Volumetric Tumour Segmentation

TL;DR

LinGuinE (Longitudinal Guidance Estimation), a PyTorch framework that combines image registration and guided segmentation to deliver lesion-level tracking and volumetric masks across all scans in a longitudinal study from a single radiologist prompt, is introduced.

Abstract

Longitudinal volumetric tumour segmentation is critical for radiotherapy planning and response assessment, yet this problem is underexplored and most methods produce single-timepoint semantic masks, lack lesion correspondence, and offer limited radiologist control. We introduce LinGuinE (Longitudinal Guidance Estimation), a PyTorch framework that combines image registration and guided segmentation to deliver lesion-level tracking and volumetric masks across all scans in a longitudinal study from a single radiologist prompt. LinGuinE is temporally direction agnostic, requires no training on longitudinal data, and allows any registration and semi-automatic segmentation algorithm to be repurposed for the task. We evaluate various combinations of registration and segmentation algorithms within the framework. LinGuinE achieves state-of-the-art segmentation and tracking performance across four datasets with a total of 456 longitudinal studies. Tumour segmentation performance shows minimal degradation with increasing temporal separation. We conduct ablation studies to determine the impact of autoregression, pathology specific finetuning, and the use of real radiologist prompts. We release our code and substantial public benchmarking for longitudinal segmentation, facilitating future research.

Paper Structure

This paper contains 13 sections, 5 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methods
  3. The LinGuinE Framework
  4. Guided Volumetric Segmentation
  5. Test Datasets
  6. Experimental Results
  7. Segmentation Performance
  8. Temporal Direction Agnosticism and Degradation
  9. Ablation Studies
  10. Radiologist Inputs
  11. Impact of prompt boosting
  12. Auto-regressive propagation
  13. Conclusion

Figures (2)

  • Figure 1: LinGuinE segmentations of two longitudinal studies in the 4DCBCT dataset. The center of the screening label was propagated to obtain segmentations (red contours) on subsequent timepoints. The ground truth is denoted by yellow contours.
  • Figure 2: LinGuinE displays minimal performance degradation with elapsed time between the baseline and follow-up. Points represent mean Dice at various timepoints with the standard error on the mean displayed as error bars. The line of best fit for each method is shown for trend comparison - LinGuinE has the smallest slope.