Abstract
Deep neural networks have demonstrated impressive accuracy in supervised learning tasks. However, their lack of transparency makes it hard for humans to trust their results, especially in safety-critical domains such as healthcare. To address this issue, recent explanation-guided learning approaches proposed to align the gradient-based attention map to image regions annotated by human experts, thereby obtaining an intrinsically human-centered model. However, the attention map these methods are based on may fail to causally attribute the model predictions, thus compromising their validity for alignment. To address this issue, we propose a novel human-centered framework based on counterfactual generation. In particular, we utilize the counterfactual generation's ability for causal attribution to introduce a novel loss called the CounterFactual Alignment (CF-Align) loss. This loss guarantees that the features attributed by the counterfactual generation for the classifier align with the human annotations. To optimize the proposed loss that entails a counterfactual generation with an implicit function form, we leverage the implicit function theorem for backpropagation. Our method is architecture-agnostic and, therefore can be applied to any neural network. We demonstrate the effectiveness of our method on a lung cancer diagnosis dataset, showcasing faithful alignment to humans.
Introduction
In the high-stakes world of medical imaging, it’s not enough for an AI to be right; it has to be right for the right reasons. A model that identifies lung cancer because it spotted a specific hospital’s watermark in the corner of a scan is a liability, not an asset. This research addresses the "shortcut learning" problem—where models rely on spurious correlations—by moving beyond simple visual associations and into the realm of causal reasoning.
The Challenge of Shortcut Learning
Deep learning models are notorious for finding shortcuts in training data that do not hold up in the real world. In radiology, this might mean a model learns to associate a certain type of scanner or a specific annotation mark with a "malignant" label, rather than analyzing the actual tumor.
Existing "alignment" methods attempt to fix this by forcing the model's attention maps (like Grad-CAM) to overlap with expert-drawn boxes. However, this is merely associational alignment. Just because a model is looking at the right area doesn't mean that area is causing the decision. Two models can have identical attention patterns but completely different internal logic. The goal of this work is to align the causal mechanism of the AI with that of the human expert.
Proposed Method: Causal Alignment via Counterfactuals
The core innovation is the Causal Alignment Loss (CF-Align), built on three main pillars:
- Counterfactual (CF) Generation: To find out what truly "causes" a model's decision, the framework asks: "What is the smallest change I can make to this image to flip the diagnosis from benign to malignant?" The areas that change in this counterfactual image are the model's true causal factors.
- Alignment via Implicit Function Theorem: Generating a counterfactual is an optimization problem. To train the main model using this process, the authors use the Implicit Function Theorem (IFT) combined with a Conjugate Gradient solver. This allows for efficient gradient calculation through the counterfactual generation step without storing massive Hessian matrices.
- Hierarchical Alignment: Radiologists follow a pipeline: Image $\to$ Attributes (e.g., "spiculation" or "margin") $ o$ Diagnosis. The proposed hierarchical architecture mimics this, ensuring the model identifies correct clinical attributes before making a final call.
Experimental Results
The researchers evaluated their method on several key medical imaging benchmarks, including LIDC-IDRI (thoracic CT) and CBIS-DDSM (mammography). To strictly test for shortcut learning, they added synthetic shortcuts (like small "+" or "−" symbols) to images. While baseline models were easily fooled, the causally aligned model maintained high accuracy by focusing on tumor margins and textures.
As shown in the results, the proposed method achieved a CAM Precision of 0.751 on the LIDC-IDRI dataset, significantly outperforming baseline methods like Ross et al. (0.034) and Zhang et al. (0.068). This indicates that the model's reasoning is far more concentrated on the actual lesion rather than background noise or shortcuts.
Conclusion
By identifying the minimal change needed to flip a label, this framework isolates the features the model actually relies on for its logic. While the computational complexity of estimating the implicit Jacobian matrix is higher than standard backpropagation, the leap in interpretability and robustness makes it a vital step toward reliable, human-centered AI in healthcare. Future work will focus on more efficient solvers and extending the framework to multi-modal data.
Summary updated