PCa-Mamba: Spatiotemporal state space models for prostate cancer detection in multi-parametric MRI

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Multiparametric MRI (mpMRI) is crucial for diagnosing clinically significant prostate cancer (csPCa).
• Dynamic contrast-enhanced MRI (DCE-MRI) offers valuable microvascular information but is underutilized due to data challenges.
• Existing deep learning methods often overlook DCE-MRI's temporal dynamics.

Purpose of the Study:

• To introduce PCa-Mamba, a novel framework leveraging state-space models (SSMs) for comprehensive mpMRI analysis in csPCa detection.
• To fully integrate DCE-MRI temporal dynamics with T2-weighted imaging (T2) and diffusion-weighted imaging (DWI) spatial data.
• To address spatial-temporal data incompatibility and improve csPCa detection rates.

Main Methods:

• Developed PCa-Mamba, a framework using two SSM modules: one for temporal dynamics (DCE-MRI) and one for spatial contrast (T2/DWI).
• Incorporated pharmacokinetic (PK) regularization with the Tofts model for temporal SSM.
• Utilized permuted sequentialization in the spatial SSM and introduced a dropout mechanism for handling missing DCE-MRI data.

Main Results:

• PCa-Mamba demonstrated superior performance in csPCa detection compared to existing models on in-house and PI-CAI datasets.
• The framework effectively integrated spatial and temporal features for enhanced spatiotemporal representations.
• The dropout mechanism enabled robust performance even without DCE-MRI data.

Conclusions:

• PCa-Mamba successfully incorporates DCE-MRI temporal dynamics into deep learning for csPCa detection.
• The proposed method improves lesion-wise diagnosis, particularly for small and peripheral zone prostate cancers.
• This framework highlights the potential of fully utilizing mpMRI data for more accurate prostate cancer assessment.