Exploring Counterfactual Alignment Loss towards Human-Centered AI

Mingzhou Liu, Xinwei Sun, Ching-Wen Lee, Yu Qiao, Yizhou WangSubmission to ICLR 2024 · Withdrawn

TL;DR

Deep learning models in medical imaging frequently rely on spurious background correlations rather than clinically meaningful features. This paper introduces CounterFactual Alignment (CF-Align), an architecture-agnostic loss function that forces a model's causal decision boundaries to align with human expert annotations. By framing counterfactual generation as an inner optimization problem and solving its gradient via the Implicit Function Theorem, the authors train classifiers that focus on actual pathology rather than background artifacts.

What Problem Does This Paper Address?

Standard deep neural networks often achieve high benchmark accuracy by exploiting confounding shortcuts. For instance, in thoracic CT scans, a model might predict lung cancer using textures from the ribs or spine rather than inspecting the pulmonary nodule itself.

Previous "explanation-guided" learning methods tried to fix this by regularizing gradient- or attention-based saliency maps toward expert bounding boxes. However, attention and gradients do not necessarily reflect causal decision-making—an input region can show high gradient magnitude without being the actual causal driver of the network’s final class prediction. The authors address the need for an ante-hoc (intrinsically explainable) training objective that ensures the features causally responsible for a model's prediction genuinely lie inside human-annotated regions of interest.

Methods

The authors ground their alignment framework in counterfactual explanations (CE). A counterfactual image $x^*(\theta)$ is defined as the minimal, sparse modification to an input image $x$ that flips the model’s prediction to a target class $y^*$:

$$x^*(\theta) := \arg\min_{x'} \text{CE}(f_\theta(x'), y^*) + \lambda d(x, x')$$

If the model is genuinely human-centered, the delta $(x^*(\theta) - x)$ should occur only within the expert-annotated region mask $r$.

Forward:  x ──► [Classifier f_θ] ──► Counterfactual x*(θ) ──► Compute L_align (penalize edits outside r)
                                                                     │
Backward: θ ◄── [Implicit Function Theorem + Conjugate Gradients] ◄──┘
  1. CF-Align Loss: The penalty explicitly measures counterfactual modifications falling outside the expert mask $r$: $$\mathcal{L}_{\text{align}}(\theta) := \frac{1}{n} \sum_{i=1}^n \|(1 - r_i) \odot (x^*(\theta) - x)\|_1$$ The full objective balances cross-entropy and alignment: $\mathcal{L}(\theta) = \mathcal{L}_{\text{cls}}(\theta) + \alpha \mathcal{L}_{\text{align}}(\theta)$.

  2. Implicit Function Theorem (IFT) Optimization: Because $x^*(\theta)$ is defined via an $\arg\min$ optimization rather than an explicit closed-form layer, standard backpropagation fails. The authors use the IFT to compute the implicit gradient $\nabla_\theta x^*(\theta)$ using the inverse Hessian of the counterfactual objective and solve it iteratively with a conjugate gradient solver (via TorchOpt).

  3. Extension to Clinical Attributes: The framework extends to intermediate attributes (e.g., nodule margin, spiculation) by decomposing the model into $g_\gamma(f_\theta(x))$ and intervening on key causal attributes identified via prior causal discovery methods.

Datasets

The authors evaluate their method on the LIDC-IDRI benchmark:

  • Task: Pulmonary nodule benign vs. malignant classification using full 2D thoracic CT slices (without cropping out the nodules).
  • Data Scale & Split: 1,213 total nodules split at the patient level into training ($n=731$), validation ($n=244$), and test ($n=238$).
  • Annotations: Physician-provided nodule bounding boxes, malignancy ratings (1–3 as benign, 4–5 as malignant), and six clinical attributes (subtlety, calcification, margin, spiculation, lobulation, texture).

Key Insights and Ideas

  • Causality over Saliency: Rather than constraining passive gradient maps (which can be noisy or deceptive), penalizing counterfactual generation forces the network's active decision mechanism to stay within the mask.
  • Bi-level Differentiable Training: By treating counterfactual generation as an inner loop optimization and solving gradients through the IFT, the model learns human alignment end-to-end without requiring specialized prototype architectures.
  • Generalization through Causal Anchoring: Restricting features to domain-relevant regions actually improves out-of-sample classification performance by preventing the model from overfitting to site-specific or contextual imaging artifacts.

Key Results

  • Saliency Map Precision: CF-Align achieved an 81% saliency precision (ratio of attention inside the true nodule box), compared to roughly 0%–50% across standard baselines (e.g., Vanilla, BagNets, ICNN, CDEP, RRR, and CIDG), which repeatedly attended to bones and lung walls.
  • Classification Accuracy: CF-Align reached ~71% accuracy, substantially outperforming baselines (mostly hovering between 50% and 56%) and matching the Oracle model (~72%), which had the unfair advantage of receiving pre-cropped nodule patches as inputs.
  • LLM Integration: By piping predicted clinical attributes and saliency regions into GPT-4 prompts, the authors demonstrated automated generation of structured, clinically faithful radiology reports.

Limitations and Future Work

  • Accuracy–Explainability Trade-off: The hyperparameter $\alpha$ must be tuned carefully; over-regularizing might discard valid microscopic image features (e.g., fine texture or contextual parenchyma) that fall outside coarse bounding boxes.
  • Computational Complexity: Computing implicit gradients through the Hessian via conjugate gradients adds non-trivial training overhead compared to standard backpropagation.
  • Scope of Evaluation: Validation was conducted primarily on a single 2D CT slice classification task; testing scalability to full 3D volumes and other safety-critical multimodal domains remains an open direction.

Summary updated

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 safe-critic 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.
BibTeX
@misc{liu2023exploring,
  author = {Mingzhou Liu and Xinwei Sun and Ching-Wen Lee and Yu QIAO and Yizhou Wang},
  title = {{Exploring Counterfactual Alignment Loss towards Human-Centered AI}},
  year = {2023},
  eprint = {2310.01766},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.01766},
}

Loading related content...

Figure