Sine Wave Normalization for Deep Learning-Based Tumor Segmentation in CT/PET Imaging

TL;DR

The paper addresses the challenge of segmentation in PET/CT amid strong inter-tracer and scanner variability. It introduces SineNormal, applying $ \mathbf{y} = \sin(\mathbf{a} \cdot \mathbf{x}_{\mathrm{PET}})$ with $ \mathbf{a}=[a_1,\ldots,a_c]$, to produce sine-transformed PET channels that emphasize metabolic variations and create concentric ring patterns. This module is integrated into a 3D nnUNet ResEnc(M)-based architecture with UMamba blocks and four-channel input (CT, PET, and two sine-transformed PET channels) after 0-1 normalization, followed by post-processing for efficiency. The approach targets the AutoPET III multitracer dataset with 1611 training cases, a dynamic sliding-window post-processing pipeline, and a plan for rigorous future validation to demonstrate generalizability.

Abstract

This report presents a normalization block for automated tumor segmentation in CT/PET scans, developed for the autoPET III Challenge. The key innovation is the introduction of the SineNormal, which applies periodic sine transformations to PET data to enhance lesion detection. By highlighting intensity variations and producing concentric ring patterns in PET highlighted regions, the model aims to improve segmentation accuracy, particularly for challenging multitracer PET datasets. The code for this project is available on GitHub (https://github.com/BBQtime/Sine-Wave-Normalization-for-Deep-Learning-Based-Tumor-Segmentation-in-CT-PET).