[Paper Review] Z-SSMNet: Zonal-aware Self-supervised Mesh Network for Prostate Cancer Detection and Diagnosis with Bi-parametric MRI

Yuan Yuan, Euijoon Ahn+3 more|arXiv (Cornell University)|Dec 12, 2022

AI in cancer detection4 citations

TL;DR

Z-SSMNet is a zonal-aware self-supervised mesh network that leverages multi-dimensional convolutions (2D/2.5D/3D) and self-supervised pre-training on large-scale unlabeled bi-parametric MRI (bpMRI) to enhance prostate cancer detection and diagnosis. It achieves state-of-the-art performance with an AP score of 0.690 and AUROC of 0.909 on the PI-CAI challenge's closed testing phase, demonstrating superior generalization and robustness on multi-center, multi-scanner data.

ABSTRACT

Bi-parametric magnetic resonance imaging (bpMRI) has become a pivotal modality in the detection and diagnosis of clinically significant prostate cancer (csPCa). Developing AI-based systems to identify csPCa using bpMRI can transform PCa management by improving efficiency and cost-effectiveness. However, current state-of-the-art methods using convolutional neural networks (CNNs) are limited in learning in-plane and three-dimensional spatial information from anisotropic images. Their performances also depend on the availability of large, diverse, and well-annotated bpMRI datasets. We propose a Zonal-aware Self-supervised Mesh Network (Z-SSMNet) that adaptively integrates multi-dimensional (2D/2.5D/3D) convolutions to learn dense intra-slice information and sparse inter-slice information of the anisotropic bpMRI in a balanced manner. A self-supervised learning (SSL) technique is proposed to pre-train our network using large-scale unlabeled data to learn the appearance, texture, and structure semantics of bpMRI. It aims to capture both intra-slice and inter-slice information during the pre-training stage. Furthermore, we constrained our network to focus on the zonal anatomical regions to further improve the detection and diagnosis capability of csPCa. We conducted extensive experiments on the PI-CAI dataset comprising 10000+ multi-center and multi-scanner data. Our Z-SSMNet excelled in both lesion-level detection (AP score of 0.633) and patient-level diagnosis (AUROC score of 0.881), securing the top position in the Open Development Phase of the PI-CAI challenge and maintained strong performance, achieving an AP score of 0.690 and an AUROC score of 0.909, and securing the second-place ranking in the Closed Testing Phase.

Motivation & Objective

• To address the limitations of CNN-based methods in capturing both intra-slice and inter-slice spatial information from anisotropic bpMRI scans.
• To reduce reliance on large, diverse, and fully annotated bpMRI datasets by leveraging self-supervised pre-training on unlabeled data.
• To improve detection and diagnosis of clinically significant prostate cancer (csPCa) by focusing the model on zonal anatomical regions of the prostate.
• To develop a unified deep learning framework that balances dense intra-slice and sparse inter-slice feature learning across 2D, 2.5D, and 3D convolutions.
• To achieve strong performance on real-world, multi-center, multi-scanner bpMRI data without extensive fine-tuning.

Proposed method

• The model employs a hybrid convolutional architecture with adaptive integration of 2D, 2.5D, and 3D convolutions to capture both local texture and global spatial context in anisotropic bpMRI.
• A self-supervised learning (SSL) objective is introduced to pre-train the network on large-scale unlabeled bpMRI data, learning appearance, texture, and structural semantics.
• Zonal attention mechanisms are applied to constrain feature learning to anatomically relevant regions of the prostate, improving localization and diagnostic relevance.
• The network is pre-trained using contrastive learning on cropped patches from multi-slice bpMRI volumes, enabling robust feature representation without annotations.
• The final model is fine-tuned end-to-end for lesion-level detection and patient-level diagnosis using weakly supervised signals.
• The architecture is designed to handle variable slice thickness and inter-slice spacing typical in clinical bpMRI acquisitions.

Experimental results

Research questions

• RQ1 Can a self-supervised pre-training strategy effectively learn meaningful representations from unlabeled bi-parametric MRI to improve downstream csPCa detection?
• RQ2 Does the integration of multi-dimensional convolutions (2D/2.5D/3D) enhance the model’s ability to capture both local and global spatial dependencies in anisotropic MRI?
• RQ3 To what extent does zonal attention improve the diagnostic performance of deep learning models in prostate cancer detection?
• RQ4 How does Z-SSMNet generalize across diverse multi-center and multi-scanner bpMRI data compared to existing CNN-based methods?
• RQ5 Can a zonal-aware, self-supervised approach achieve state-of-the-art performance without relying on large-scale annotated datasets?

Key findings

• Z-SSMNet achieved an AP score of 0.690 and AUROC of 0.909 in the closed testing phase of the PI-CAI challenge, securing second place.
• In the open development phase, it achieved an AP score of 0.633 and AUROC of 0.881, demonstrating strong generalization on unseen data.
• The self-supervised pre-training stage effectively captured appearance, texture, and structural semantics from unlabeled bpMRI data, improving downstream performance.
• The integration of zonal attention significantly enhanced lesion localization and diagnostic accuracy by focusing on anatomically relevant regions.
• The model outperformed existing CNN-based methods in handling the anisotropic nature of bpMRI by balancing dense intra-slice and sparse inter-slice feature learning.
• The architecture demonstrated robustness across diverse multi-center and multi-scanner bpMRI datasets, confirming its clinical applicability.