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

TL;DR

The Zonal-aware Self-supervised Mesh Network (Z-SSMNet) is proposed, which adaptively integrates multi-dimensional convolutions to learn dense intra-slice information and sparse inter-slice information of the anisotropic bpMRI in a balanced manner and constrain the network to focus on the zonal anatomical regions to improve the detection and diagnosis capability of csPCa.

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.

Figures (7)

• Figure 1: A schematic of our Z-SSMNet. The one-hot encoded zonal masks (CG, PZ, and BG (background)) are concatenated with the bpMRI images (T2W, DWI, and ADC) as inputs to provide anatomical priors to the model. Our SSL pre-training involves recovering the original sub-volumes of images from their corresponding corrupted images. The trained network is then fine-tuned for accurate detection and diagnosis of csPCa.
• Figure 2: The network architecture of Z-SSMNet. The mesh structure concurrently combines numerous representation processes in a latent manner to dynamically create a balanced representation through adaptive learning for anisotropic information across axes. Within basic modules, both multi-dimensional and multi-level features are latently fused to harness the benefits of both 2D and 3D representations, enabling more precise modelling for target regions. Supervisory information is conveyed to six extra output branches, ensuring a thorough training of shallow layers.
• Figure 3: Visualization of csPCa detection maps from our proposed Z-SSMNet and comparing methods. The detection maps and corresponding T2W, DWI, and ADC scans (columns) are shown for three cases (rows). All the methods were trained on Set 2 (semi-supervised setting). The numbers corresponding to the bottom of the images represent the maximum probability value of the detection map.
• Figure 4: Overview of the zonal mask generation process. The T2W and ADC images are used for model training and testing.
• Figure 5: The examples of the generated zonal masks with and without post-processing.