Lesion-guided selective multi-modal integration for prostate cancer segmentation and PI-RADS grading in MP-MRI

Prostate cancer (PCa) presents a significant global health challenge affecting men. Accurate segmentation and grading of PCa lesions in multiparametric Magnetic Resonance Imaging (mp-MRI) are essential for effective diagnosis and treatment planning.

This study aimed to develop and validate an automated model for PCa lesion segmentation and Prostate Imaging Reporting and Data System (PI-RADS) grading in mp-MRI.

The lesion's perceived characteristics are strongly related to both imaging modalities and lesion locations. Therefore, we propose a Lesion-guided Selective Multi-modal Integration (LeSMI) module. This module incorporates two advanced mechanisms-Dynamic Modality Weighting (DMW) and Localized Lesion Attention (LLA)-to dynamically integrate crucial information across and within imaging modalities. Specifically, DMW operates on the mp-MRI inputs (T2-weighted (T2w) images, diffusion-weighted imaging (DWI), and apparent diffusion coefficient (ADC) maps) to dynamically assign weights to each modality, thereby integrating complementary information and enhancing feature identification across different contexts. LLA, on the other hand, maintains spatial structure information within each modality for precise lesion localization. Inspired by clinical workflows, our framework is employed through a two-stage Prostate Cancer Segmentation and Grading (PCaSG) strategy, leveraging knowledge from segmentation to improve PI-RADS grading performance. We validated our method using two publicly available datasets, namely, Prostate158 and PI-CAI Challenge, to assess its advantages over other methods. For the Prostate158 dataset, we used the officially reported partition with 119 cases for training, 20 for validation, and 19 for testing. In contrast, the PI-CAI Challenge dataset, which lacks predefined splits, was randomly divided into 180 for training, 20 for validation, and 20 for testing. In addition to these dataset partitions, 5-fold cross-validation was conducted on both the Prostate158 and PI-CAI Challenge datasets to provide a more robust and comprehensive statistical evaluation of the model's performance.

Evaluated on the Prostate158 and PI-CAI Challenge datasets, our method demonstrated superior performance, achieving a Dice Similarity Coefficient (DSC) of 51.30% and a lesion-level quadratic-weighted kappa score ( $QW{K}_l$ ) of 62.48% on Prostate158, and a DSC of 43.81% and a $QW{K}_l$ of 42.98% on PI-CAI. These results represent improvements of up to 2% in DSC and 17% in $QW{K}_l$ over current state-of-the-art models on Prostate158, and enhancements of 4% in DSC and 3% in $QW{K}_l$ on PI-CAI.

The proposed model's robustness in handling diverse lesion presentations, combined with its reliable assessments, underscores its significant clinical applicability. Our model offers substantial advancements in both segmentation accuracy and PI-RADS grading, addressing the challenges of inter-reader variability and the need for high expertise in conventional diagnostic practices. This technological innovation holds promise for enhancing early, accurate detection and risk assessment in prostate cancer management, ultimately improving patient outcomes.