Mask Enhanced Deeply Supervised Prostate Cancer Detection on B-mode Micro-Ultrasound

TL;DR

Prostate cancer targeting on micro-ultrasound is challenged by subtle tumor features and artifacts. The authors propose MedMusNet, a 3D UNet-based architecture with a mask-enhanced module and multi-scale deep supervision to automatically segment clinically significant cancer on B-mode micro-US. On a prospective cohort of 64 patients with MRI fusion biopsy ground truth, MedMusNet achieves superior pixel-level segmentation (notably a higher Dice score than baselines) and competitive lesion- and patient-level performance, approaching expert readers while improving specificity and accuracy in key comparisons. The study introduces a practical pipeline to map MRI-based labels to micro-US, demonstrates potential for guiding ultrasound-only biopsy decisions, and outlines limitations including single-site data and false-positive rates that warrant further validation.

Abstract

Prostate cancer is a leading cause of cancer-related deaths among men. The recent development of high frequency, micro-ultrasound imaging offers improved resolution compared to conventional ultrasound and potentially a better ability to differentiate clinically significant cancer from normal tissue. However, the features of prostate cancer remain subtle, with ambiguous borders with normal tissue and large variations in appearance, making it challenging for both machine learning and humans to localize it on micro-ultrasound images. We propose a novel Mask Enhanced Deeply-supervised Micro-US network, termed MedMusNet, to automatically and more accurately segment prostate cancer to be used as potential targets for biopsy procedures. MedMusNet leverages predicted masks of prostate cancer to enforce the learned features layer-wisely within the network, reducing the influence of noise and improving overall consistency across frames. MedMusNet successfully detected 76% of clinically significant cancer with a Dice Similarity Coefficient of 0.365, significantly outperforming the baseline Swin-M2F in specificity and accuracy (Wilcoxon test, Bonferroni correction, p-value<0.05). While the lesion-level and patient-level analyses showed improved performance compared to human experts and different baseline, the improvements did not reach statistical significance, likely on account of the small cohort. We have presented a novel approach to automatically detect and segment clinically significant prostate cancer on B-mode micro-ultrasound images. Our MedMusNet model outperformed other models, surpassing even human experts. These preliminary results suggest the potential for aiding urologists in prostate cancer diagnosis via biopsy and treatment decision-making.

Figures (8)

• Figure 1: (Top) Examples of micro-ultrasound prostate images of three patients, (Bottom) depicting the prostate boundary (blue) and clinically significant cancer (purple).
• Figure 2: Pre-processing pipeline to facilitate the labeling of prostate cancer from MRI onto micro-ultrasound. The steps involve the reconstruction of a Cartesian 3D micro-ultrasound volume, to which the MRI is registered allowing the projection of labels from MRI onto micro-ultrasound in the native pseudo sagittal space.
• Figure 3: Illustration of micro-ultrasound prostate scanning. The transrectal micro-ultrasound probe rotates from left to right, generating a series of 2D images positioned in the pseudo-sagittal oblique plane, separated by the rotation angle $\theta$. Yellow indicates the transition zone, blue represents the peripheral zone, and purple denotes cancer lesions.
• Figure 4: Illustration of mapping the location of clinically significant cancer (ISUP Grade Group $\geq 2$) on micro-ultrasound images using MR images. (a) The MRI is manually registered using affine transformations to the 3D Cartesian micro-ultrasound scan using 3D Slicer guided by the prostate boundary (Blue). The radiologists' annotations of biopsy-confirmed MRI-visible lesions (purple) are projected from MRI onto 3D Cartesian B-mode micro-ultrasound images. (b) The ground truth cancer labels (purple) are projected back onto the native pseudo-sagittal images, and refined by export in the native space to reduce the effect of registration errors and interpolation artifacts.
• Figure 5: Overview of the proposed MedMusNet model. Based on a 3D UNet-like backbone, the model incorporates the spatial relationship explicitly and enhanced global features captured by mask-enhanced modules and the multi-scale deep supervision. $\mathcal{L}_n$ represents multi-scale losses and the overall loss $\mathcal{L}$ is their weighted sum (Eq. \ref{['eq:Ln']} and Eq. \ref{['eq:L']}).