<scp>CITMambaNet</scp> : Cross Modality Progressive Fusion <scp>ITMamba</scp> Network for <scp>PET</scp> ‐ <scp>CT</scp> Medical Image Segmentation

ABSTRACT Positron Emission Tomography (PET) and Computed Tomography (CT) can provide rich metabolic information and anatomical background, respectively. The combination of PET and CT can fully leverage their advantages through modality fusion. PET‐CT image segmentation can provide an accurate and precise representation of organs within the body. Recently, deep learning has been a research hotspot in medical image segmentation. This paper develops Cross Modality Progressive Fusion ITMambaNet (CITMambaNet) for tumor segmentation of both 3D and 2D PET‐CT images. It is based on the dual UNETR framework. Firstly, an InceptionV2‐ToMamba block (ITMamba) is introduced for model long‐range dependency, which has fewer parameters than transformers by processing data sequentially through a State Space Model (SSM). Secondly, a Cross Modality Progressive Fusion block (CMPF) is designed to fuse complementary features from PET and CT. Finally, a Skip‐and‐Connect‐next block (SC‐next) is used to effectively combine semantic information from the encoder and decoder. The proposed CITMambaNet with 3D and 2D versions was validated on the Hecktor 2021 and Soft‐tissue‐Sarcoma (STS) datasets, with Dice coefficients of 74.56% and 75.73%, and average HD95 values of 8.47 mm and 2.53 mm, respectively. The experimental results show that our model achieved the best performance compared with state‐of‐the‐art methods.