Tumour Segmentation and Volume estimation from PET Lymphoma Data Using a Pre-trained U-Net Model and Probability Contour Framework

Abstract

Accurate tumour delineation is critical for utilizing total metabolic tumour volume (TMTV) as a reproducible prognostic biomarker in lymphoma PET imaging [1]. This study focuses on achieving robust tumour segmentation specifically from PET data [1] by applying a pre-trained U-Net model [2] and a kernel-smoothed probability contour framework [3]. Additionally, ongoing segmentation experiments on various other PET datasets are underway to validate and generalize this approach further

Tumour segmentation was conducted using a pre-trained U-Net model, initially developed on the AutoPET II PET challenge dataset [4]. The model achieved a Dice similarity coefficient of 0.65 on the new lymphoma PET benchmark dataset [1]. To enhance interpretability and reliability, a kernel-smoothed probability contour framework was applied as a post-processing step, generating voxel-wise uncertainty information for each segmented subvolume without compromising Dice accuracy.

The U-Net model effectively segmented lymphoma tumours across multiple PET datasets, consistently demonstrating stable and reproducible performance. The probability contour framework successfully provided uncertainty quantification, highlighting regions of high and low confidence in PET segmentation results. This combined methodology presents a robust, interpretable solution for tumour delineation specifically optimized for PET lymphoma imaging.

Integrating a pre-trained U-Net model with a kernel-smoothed probability contour approach enables standardized and reproducible tumour segmentation from PET data in lymphoma. The provision of uncertainty metrics significantly enhances the segmentation's reliability, supporting the potential of TMTV as a robust clinical biomarker. Future research leveraging this PET-focused framework will continue to explore and validate its clinical utility across additional PET lymphoma datasets.

Please click on the 'PDF' for the full abstract!