A PET/CT Cross-Modal Wavelet Fusion and Pseudo-Mask Guided Network With Frozen SAM Decoder for Multiple Myeloma Segmentation.

Abstract

Accurate segmentation of Multiple Myeloma (MM) lesions from PET/CT is critical for prognosis but challenging due to lesion heterogeneity and cross-modal frequency discrepancies. This study aims to develop a robust, parameter-efficient framework to address these complexities. We introduce a Cross-Modal Dual-Wavelet Fusion Network with a Frozen SAM Decoder. Unlike standard adaptations, we replace the heavy SAM image encoder with a custom dual-branch wavelet encoder. This design explicitly aligns high-frequency anatomical edges from CT with low-frequency metabolic cues from PET via a "decompose-and-inject" mechanism. Furthermore, we freeze the pre-trained SAM mask decoder to leverage robust geometric priors and employ a lightweight Tiny Pseudo Decoder for auxiliary boundary supervision. Extensive experiments on an in-house MM dataset (N=161) and the public HECKTOR 2022 dataset demonstrate that our method achieves Dice scores of 0.8323 and 0.8465, respectively, significantly outperforming state-of-the-art baselines including nnU-Net and MedSAM ($p < 0.05$). Notably, the model requires only 10.28M parameters-a $\approx$96% reduction compared to the standard SAM (271.24 M)-while maintaining superior accuracy. The proposed framework effectively bridges the gap between lightweight deployment and high-performance cross-modal segmentation. The approach offers a practical, lightweight alternative for PET/CT segmentation, facilitating deployment in clinical workflows by combining strong accuracy, cross-dataset generalizability, and markedly improved computational efficiency. The model code is available at https://github.com/HanXinfun/DualwaveSAM.