Segmentation of Prostate Tumour Volumes from PET Images is a Different Ball Game

TL;DR

This work tackles the challenge of segmenting intra-prostatic gross tumour volumes (GTV) from PSMA-PET images, where standard normalisation from CT/MR struggles. It introduces Feature Clipping Normalisation (FCN), a PET-specific preprocessing that clips intensities using thresholds derived from physician delineations (e.g., $30\%$–$40\%$ of $SUV_{max}$), and evaluates this with four U-Net variants within the nnU-Net framework across two tracers. FCN consistently improves segmentation performance over standard normalisations, with the best results obtained by the IB-Attention-U-Net on the 18-F dataset (e.g., $NSD\approx0.761$, $DSC\approx0.840$, $HD_{95}\approx3.354$). These findings support PET-tailored preprocessing and robust network architectures for accurate, reproducible PSMA-PET GTV delineation, with open-source integration and implications for radiotherapy planning.

Abstract

Accurate segmentation of prostate tumours from PET images presents a formidable challenge in medical image analysis. Despite considerable work and improvement in delineating organs from CT and MR modalities, the existing standards do not transfer well and produce quality results in PET related tasks. Particularly, contemporary methods fail to accurately consider the intensity-based scaling applied by the physicians during manual annotation of tumour contours. In this paper, we observe that the prostate-localised uptake threshold ranges are beneficial for suppressing outliers. Therefore, we utilize the intensity threshold values, to implement a new custom-feature-clipping normalisation technique. We evaluate multiple, established U-Net variants under different normalisation schemes, using the nnU-Net framework. All models were trained and tested on multiple datasets, obtained with two radioactive tracers: [68-Ga]Ga-PSMA-11 and [18-F]PSMA-1007. Our results show that the U-Net models achieve much better performance when the PET scans are preprocessed with our novel clipping technique.

Figures (3)

• Figure 1: Example of CT (left) and [$^{68}$Ga]Ga-PSMA-11 PET (right) images with annotated prostate gland (red) and tumour (green). Center: Cropped pelvic regions with the prostate and tumour (top: CT and bottom: PET). These images highlight the importance of correct modality for RT, as the tumour volume is more pronounced in PET than in CT.
• Figure 2: Left and Center: Performance of semi-automatic contouring approaches on the intra-prostatic SUV, of the PET images for 68-Ga and 18-F tracers. All voxels with uptake values equal-to and above the SUVmax$\%$ threshold, are considered as tumour, and the rest as background. For Dice coefficient (DSC) and normalised surface dice (NSD), the higher their value the better the prediction, whereas, for the Hausdorff Distance 95% percentile (HD-95) is the other way around. The best performance for all three metrics is achieved between 30%-40%. Right: The absolute threshold uptake values of all images for a given percent. Most of the average threshold values lie in the range 3-10.
• Figure 3: A qualitative accuracy comparison of U-Net and Attention U-Net and their IB extended variants on the prostate tumour segmentation task. All the PET images were scaled using the FCN algorithm, and then trained using the U-Net models. As one can observe, the IB-versions perform better (fewer instances of false positives), with the IB-Attention-Net, achieving the best performance. Furthermore, the results show that the tumours are prominently differentiable from the background with 18-F tracer, in contrast to 68-Ga.