MSTM-Net: a two-stage prostate cancer segmentation network based on swin-transformer-mamba architecture

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Magnetic resonance imaging (MRI) is crucial for prostate cancer diagnosis.
• Accurate segmentation of prostate cancer lesions is vital for staging and treatment planning.

Purpose of the Study:

• To develop an automated two-stage segmentation framework for prostate cancer detection using multimodal MRI.
• To improve the accuracy and efficiency of prostate cancer segmentation.

Main Methods:

• A two-stage segmentation framework was proposed, first segmenting the prostate gland and then performing lesion segmentation within the region of interest.
• A novel MSTM-Net architecture incorporating a Swin Transformer decoder, Mamba module for state-space modeling, and multi-scale feature fusion was developed.
• T2-weighted MRI and apparent diffusion coefficient (ADC) maps were used as input, spatially aligned and concatenated.

Main Results:

• The MSTM-Net achieved a Dice score of 95.38% for prostate gland segmentation and 63.89% for lesion segmentation on the PROSTATEx dataset.
• The proposed method outperformed comparative networks by approximately 4% in lesion segmentation.
• Cross-dataset validation on the PI-CAI dataset showed a Dice score of 63.14%, demonstrating good generalization.

Conclusions:

• The MSTM-Net demonstrates superior performance for prostate cancer segmentation in multimodal MRI.
• The two-stage framework, multimodal fusion, and state-space modeling offer a promising approach for automated segmentation.
• Further validation on diverse datasets is recommended to enhance robustness and generalization.