Introduction
Deep learning models have shown remarkable performance in medical imaging, yet their "black box" nature remains a significant barrier to clinical adoption. To bridge this gap, researchers employ path attribution methods like Integrated Gradients (IG) to explain model predictions. These methods quantify the importance of input features by integrating gradients along a path from a "baseline" image—intended to represent the absence of features—to the actual input.
However, the choice of baseline is critical and often overlooked. In clinical practice, there is no universal "neutral" state. For instance, in a Chest X-ray, a pixel value of zero represents air, a clinically relevant feature, rather than "nothingness." Similarly, in esophageal manometry, low pressure is a diagnostic indicator of disease. Standard baselines like all-zeros, blurred images, or random noise lack medical semantic meaning, often resulting in "noisy" attribution maps that highlight irrelevant anatomical structures or artifacts, thereby undermining clinician trust.
Methodology: Medically Meaningful Baselines
To address these limitations, this work introduces a framework for generating personalized, "medically normal" counterfactual baselines. The core idea is to create a version of the patient's own scan that has been "healed" to represent a healthy state, ensuring that the attribution path focuses strictly on pathological changes.
The authors highlight that path attribution methods are essentially "blind" to the semantic meaning of the baseline. By using a counterfactual, every pixel change becomes clinically relevant. The process involves three main stages:
- Generative Modeling: A Variational Autoencoder (VAE) is trained on a large dataset of healthy medical images to learn the underlying distribution of "normal" anatomy.
- Latent Optimization: For a given pathological input, its representation in the VAE's latent space is found and iteratively optimized via gradient descent until the decoded image is classified as "healthy" by the target model.
- Path Attribution: Integrated Gradients (IG) or Expected Gradients (EG) are then computed using this generated counterfactual as the starting point.
By using a baseline that shares the same patient-specific anatomy (e.g., bone structure, positioning), the resulting attribution map cancels out healthy anatomical variance and highlights only the specific features—whether "missing" or "excessive"—that define the pathology.
Experimental Setup and Datasets
The proposed method was evaluated across three distinct medical modalities to demonstrate its generalizability:
- High-Resolution Manometry (HRM): Spatiotemporal pressure maps used for diagnosing swallowing disorders. This modality is particularly "missingness-heavy," as the absence of specific pressure patterns often indicates disease.
- Chest X-ray: Focusing on the classification and localization of pneumothorax (collapsed lung).
- Brain MRI: Applied to tumor classification and localization tasks.
Performance was measured using ground-truth segmentation masks to quantify how accurately the attribution maps localized the actual medical conditions.
Results and Discussion
The counterfactual-guided approach (CF and EG-CF) consistently outperformed traditional baselines across both qualitative and quantitative assessments.
Localization and Cleanliness
In the Manometry dataset, the counterfactual baseline significantly reduced the False Positive Attribution Rate (FPAR). Unlike standard baselines that produced noisy signals across the entire pressure map, the CF-based maps were highly focused on the regions of interest, providing a much cleaner explanation for clinicians.
Feature Importance through Ablation
The researchers conducted "mass-center ablation" tests, where pixels were removed starting from the center of the attribution map. For the counterfactual methods, the classifier's confidence dropped much more rapidly compared to other baselines. This indicates that the CF-guided attributions were more effective at identifying the features most influential to the model's decision.
Quantitative Superiority
Quantitative metrics, including Mean Squared Error (MSE) relative to ground-truth masks, confirmed the superiority of the counterfactual approach in Manometry and Brain MRI. In Chest X-rays, the method performed comparably to the best traditional baseline (Blurred), which is notable given the higher anatomical complexity and variance in X-ray images.
Conclusion and Future Work
This research highlights that the effectiveness of path attribution methods in medicine is deeply tied to the semantic relevance of the chosen baseline. By shifting from generic "blank" baselines to personalized counterfactuals, we can produce explanations that are more aligned with clinical reasoning and anatomical reality.
While the approach introduces a computational overhead—requiring approximately 2.5 seconds per image for optimization on modern hardware—this is often acceptable in clinical workflows where accuracy and interpretability are paramount. Future work will explore more advanced generative architectures, such as Diffusion Models, to further improve the fidelity of counterfactual generation in complex anatomical settings.
Summary updated