ProjectedEx: Enhancing Generation in Explainable AI for Prostate Cancer

TL;DR

ProjectedEx tackles the explainability gap in MRI-based prostate cancer diagnosis by introducing a generative framework that yields interpretable, multi-attribute explanations tied to classifier decisions. It enhances the encoder with a Feature Pyramid Encoder and uses differentiable random projections to achieve robust, multiscale representations, with StyleSpace coordinates encoding classifier-relevant attributes as affine transformations of outputs. Empirical results on the PI-CAI dataset show state-of-the-art generative quality (FID as low as $108.63$) and strong classification accuracy (up to $83.97\%$), along with demonstrable interpretability through attribute manipulation across $DWI$, $T2WI$, and $ADC$ modalities. The work promises improved clinical adoption by enabling scenario testing and transparent AI-assisted decision-making in prostate cancer management.

Abstract

Prostate cancer, a growing global health concern, necessitates precise diagnostic tools, with Magnetic Resonance Imaging (MRI) offering high-resolution soft tissue imaging that significantly enhances diagnostic accuracy. Recent advancements in explainable AI and representation learning have significantly improved prostate cancer diagnosis by enabling automated and precise lesion classification. However, existing explainable AI methods, particularly those based on frameworks like generative adversarial networks (GANs), are predominantly developed for natural image generation, and their application to medical imaging often leads to suboptimal performance due to the unique characteristics and complexity of medical image. To address these challenges, our paper introduces three key contributions. First, we propose ProjectedEx, a generative framework that provides interpretable, multi-attribute explanations, effectively linking medical image features to classifier decisions. Second, we enhance the encoder module by incorporating feature pyramids, which enables multiscale feedback to refine the latent space and improves the quality of generated explanations. Additionally, we conduct comprehensive experiments on both the generator and classifier, demonstrating the clinical relevance and effectiveness of ProjectedEx in enhancing interpretability and supporting the adoption of AI in medical settings. Code will be released at https://github.com/Richardqiyi/ProjectedEx

Figures (6)

• Figure 1: Visualization of three MRI modalities (T2WI, DWI, ADC) and corresponding prostate cancer locations. The first row shows the original images for each modality, while the second row highlights the prostate cancer regions in red.
• Figure 2: The architecture diagram illustrating the interactions between the encoder ($E$), generator ($G$), and classifier ($C$). Specifically, $D_1, D_2, D_3, D_4$ are concatenated to form $E(x)$, and $x'$ is generated as $G(E(x), C(x))$. The reconstruction loss $L_{\text{rec}}^x$ is calculated between $x$ and $x'$, while the classification loss $L_{\text{cls}}$ is computed between $C(x)$ and $C(G(x))$.
• Figure 3: A feature pyramid encoder extracts multiscale features from four layers ($L_1$ to $L_4$), processed by lightweight discriminators ($D_1$ to $D_4$) with spectral normalization. Outputs are unified at a fixed resolution of $128$ for consistent feedback across scales.
• Figure 4: Visualization of the effect of attributes on classifier logits (DWI).
• Figure 5: Visualization of the effect of attributes on classifier logits (T2WI).