Anomaly-Driven Approach for Enhanced Prostate Cancer Segmentation

TL;DR

This work tackles the challenge of clinically significant prostate cancer segmentation on biparametric MRI by introducing an anomaly-driven U-Net (adU-Net) that integrates anomaly maps, derived from reconstruction models, as additional input channels. It conducts a comparative study of four anomaly-detection architectures and demonstrates that ADC-based and multi-modal anomaly maps improve segmentation performance and generalization, particularly on external data. The findings highlight the value of incorporating deviations from healthy tissue to guide csPCa delineation and point to future directions in multi-modal fusion and anomaly-detection refinement. Overall, adU-Net shows robust gains over a strong baseline (nnU-Net), suggesting practical potential for more reliable automated csPCa identification across diverse datasets.

Abstract

Magnetic Resonance Imaging (MRI) plays an important role in identifying clinically significant prostate cancer (csPCa), yet automated methods face challenges such as data imbalance, variable tumor sizes, and a lack of annotated data. This study introduces Anomaly-Driven U-Net (adU-Net), which incorporates anomaly maps derived from biparametric MRI sequences into a deep learning-based segmentation framework to improve csPCa identification. We conduct a comparative analysis of anomaly detection methods and evaluate the integration of anomaly maps into the segmentation pipeline. Anomaly maps, generated using Fixed-Point GAN reconstruction, highlight deviations from normal prostate tissue, guiding the segmentation model to potential cancerous regions. We compare the performance by using the average score, computed as the mean of the AUROC and Average Precision (AP). On the external test set, adU-Net achieves the best average score of 0.618, outperforming the baseline nnU-Net model (0.605). The results demonstrate that incorporating anomaly detection into segmentation improves generalization and performance, particularly with ADC-based anomaly maps, offering a promising direction for automated csPCa identification.

Figures (4)

• Figure 1: Illustration of the proposed approach. The original MRI image is fed into an anomaly reconstruction network, which generates reconstructed images. By calculating the absolute difference between the original and reconstructed images, anomaly maps are obtained. These anomaly maps, along with the original images, are then fed as input to the adU-Net model, which generates a segmentation map for csPCa detection.
• Figure 2: Overview of the preprocessing pipeline used to improve prostate reconstruction. The pipeline begins with the original image, followed by the identification of a bounding box around the prostate. The image is then cropped to this region of interest, reducing irrelevant background information. Finally, a prostate mask is applied to isolate the prostate and remove surrounding structures, ensuring that the reconstruction models focus on the most relevant anatomical region.
• Figure 3: Comparison of reconstruction performance across different architectures. The vanilla and spatial autoencoders struggle with fine structural details, while the Diffusion Model and Fixed-Point GAN generate more accurate and high-resolution reconstructions.
• Figure 4: Visualization of anomaly maps for four patients. The first column shows the original MRI sequences (T2W, ADC, and DWI). The second and third columns display reconstructions from FP-GAN and the diffusion model. The fourth and fifth columns show the corresponding anomaly maps, while the final column presents the lesion masks highlighting clinically annotated csPCa regions.