Abstract
Automated prostate cancer grading from MRI is often hindered by limited labeled datasets and performance degradation across different imaging centers (domain shift). This study introduces a causality-driven framework for convolutional neural networks (CNNs) that leverages weak causal signals between feature maps to improve generalization. By integrating a Causality Extractor Module into a ResNet18 backbone and utilizing metric-based meta-learning, the approach achieves enhanced performance in one-shot learning scenarios. Experimental results on the PI-CAI dataset demonstrate that focusing on causal features rather than simple correlations leads to more robust classification and better lesion localization.
Introduction
Deep learning models in medical imaging typically act as "correlation machines," identifying features that co-occur with specific labels. However, these models often struggle to distinguish between features that causally relate to a diagnosis and spurious correlations that may vary between MRI vendors (e.g., Siemens vs. Philips). This lack of robustness is exacerbated in one-shot learning, where models must generalize from a single labeled example. This paper addresses these hurdles by bridging the gap between causality and computer vision to create more robust and interpretable diagnostic tools.
Methodology
The proposed architecture, a Causality-Driven ResNet18, incorporates a novel Causality Extractor Module designed to identify and weight features based on their "causal disposition."
The module operates through several key steps:
- Feature Interaction: It treats feature maps from the final convolutional layer as probabilities of pattern presence.
- Causal Disposition Estimation: It assesses if feature A "causes" feature B by observing the impact on B when A is removed.
- Causality Weighting: Weights are assigned to feature maps based on their role as "causes" versus "effects."
- Integration: Two variants,
mulcat(scaling by importance) andmulcatbool(binary filtering), are used to refine the feature maps.
To handle data scarcity, the framework employs Meta-DeepBDC, a meta-learning approach that uses Brownian Distance Covariance to measure similarity between support and query images.
Experimental Setup
The study utilized the PI-CAI challenge dataset, comprising over 1,500 multi-parametric MRI (mpMRI) acquisitions. To rigorously test for domain shift, the model was trained on data from SIEMENS scanners and evaluated on data from PHILIPS scanners. Labels included both Gleason Scores (GS) and International Society of Urological Pathology (ISUP) grades. A "granularity trick" was employed, where training on fine-grained GS tasks improved performance on coarser ISUP test tasks.
Results and Discussion
The causality-driven models consistently outperformed the baseline non-causal ResNet18 across various settings.
Key findings include:
- Improved Generalization: The
mulcatboolvariant achieved an AUROC of 0.614 in 4-way ISUP grading, surpassing the 0.585 baseline. - Robustness to Domain Shift: In a critical 2-way classification task (ISUP Grade 2 vs. others), the model reached an AUROC of 0.713, a 13% improvement over the baseline.
- Interpretability: Qualitative analysis indicates that the causality-driven approach produces cleaner activations focused on the actual lesion, effectively ignoring background noise.
- Ablation Study: Replacing learned causal weights with random weights led to significant performance drops, confirming the value of the extracted causal signals.
Conclusion
This work demonstrates that even "weak" causal signals derived from statistical co-occurrence can significantly boost the performance and robustness of CNNs in medical imaging. While current AUROC levels (0.6–0.7) require further improvement for autonomous clinical deployment, the causality-driven approach offers a promising path toward more reliable and human-like reasoning in AI-assisted diagnosis.
Summary updated