Explanation-Aware Learning for Enhanced Interpretability in Biomedical Imaging

Abstract: Deep neural networks for medical image diagnosis often achieve high predictive accuracy while relying on spurious or clinically irrelevant visual cues, limiting their trustworthiness in practice. Post-hoc explanation methods are widely used to visualize model decisions in the form of saliency maps; however, these explanations do not influence how models learn during training, allowing non-causal or confounding features to persist. This motivates the incorporation of explanation supervision directly into the training objective to guide model attention toward clinically meaningful regions and promote clinically grounded decision-making. This paper presents a systematic approach to integrate explanation loss into model training and analyzes how different explanation loss designs and supervision strengths influence both predictive performance and spatial faithfulness of explanations. To quantitatively assess interpretability, two complementary explanation performance metrics-annotation coverage and saliency precision-are introduced, enabling rigorous evaluation beyond qualitative visualization. Our experimental results reveal a clear trade-off between explanation quality and explanation loss coefficients. Furthermore, quantitative statistical analysis yields consistently improved explanation alignment while maintaining comparable accuracy. Experiments were conducted on annotated chest X-ray datasets; however, the proposed framework is applicable to a broad range of annotated biomedical imaging modalities. Overall, these findings demonstrate that explanation supervision is not a monolithic design choice and provide practical guidance for incorporating explanation loss into training objectives under noisy clinical annotations.

• The paper introduces explanation-aware training by integrating Grad-CAM-based explanation loss with standard BCE to guide model attention using expert annotations.
• It shows that logit-based explanation losses yield more precise and stable saliency maps, achieving a clear trade-off between annotation coverage and precision.
• The study demonstrates that explanation supervision acts as a regularizer, improving interpretability in chest X-rays with minimal impact on predictive performance.

Explanation-Aware Learning for Enhanced Interpretability in Biomedical Imaging

Introduction

Deep neural models have advanced medical imaging diagnosis across modalities, yet their deployment is impeded by a lack of interpretability and the potential to exploit spurious or clinically irrelevant features. While post-hoc methods such as Grad-CAM provide visual justifications, they do not influence the internal feature learning or model attention during training. This paper presents a systematic exploration of explanation-aware training, wherein explanation supervision is integrated directly into the model optimization process. The study offers a principled analysis of loss formulation and coefficient design, alongside quantitative interpretability evaluation, specifically in the context of chest X-ray classification with coarse radiologist annotations.

Explanation-Aware Training Framework

The proposed methodology augments standard binary cross-entropy (BCE) loss with an explanation-based term, where the coefficient governs the trade-off between predictive and explanatory objectives. The framework leverages Grad-CAM to provide per-sample saliency maps, incorporating expert-provided bounding box annotations as spatial supervision targets. Multiple explanation score formulations are considered, including various logit- and probability-based contrastive differences, reflecting the impact of both raw model outputs and class-probabilities on attention guidance.

The total loss is given as:

$L_{total} = L_{bce} + \alpha L_{exp}$

where $L_{exp}$ penalizes the misalignment between the most salient heatmap regions (via Grad-CAM) and the expert-annotated areas in each positive disease sample. Saliency thresholding and normalization approaches are adopted to mitigate the effect of noise and annotation coarseness, focusing explanation alignment on the top $k$% of activations.

Quantitative Interpretability Evaluation

To move beyond qualitative visualization, the study introduces two complementary quantitative metrics:

1. Annotation Coverage: Measures whether any salient region intersects a disease annotation, reflecting coverage under spatially coarse supervision.
2. Saliency Precision: Comprises "top saliency precision" (fraction of most salient pixels within annotations) and "all saliency precision" (fraction of total saliency mass within annotations), capturing both focused and diffuse saliency behaviors.

These metrics provide a nuanced view of explanation spatial faithfulness, penalizing models that achieve high coverage via diffuse, uninformative activations.

Empirical Evaluation

The framework is evaluated on the VinDr-CXR dataset, focusing on seven key thoracic conditions. DenseNet-121, pre-trained and fully fine-tuned for binary classification on each disease, serves as the backbone model. Models are exhaustively trained across various explanation loss types and strengths, yielding a robust distributional analysis.

Training Dynamics and Performance

Explanation-aware training is shown not to destabilize optimization or significantly degrade predictive performance. The addition of explanation supervision acts as a regularizer but maintains comparable validation accuracy to standard BCE-only models, with observed reductions in peak accuracy less than 1% and compensated by marked gains in interpretability.

Trade-off between Coverage and Precision

Increasing the explanation loss coefficient induces a clear trade-off: lower values yield high annotation coverage via spatially diffuse explanations, while higher coefficients result in saliency that is more sharply focused within annotated areas, at the cost of reduced coverage. Importantly, median classification accuracy remains stable across coefficient settings, emphasizing that interpretability improvements are not fundamentally antagonistic to predictive objectives.

Loss Formulation Effects

Logit-based explanation losses, particularly the squared logit difference, consistently produce more precise and stable attribution maps compared to probability-based or single-class logit objectives. The superiority of logit-based supervision is attributed to the smoother gradient properties and stronger implicit contrast between positive and negative predictions, while probability outputs tend to saturate and diminish explanatory gradient magnitude.

Implications and Future Directions

The findings demonstrate that explanation-aware supervision is not a uniform design choice; both the form of the explanatory signal and the strength of its incorporation into training exert heterogeneous effects on interpretability metrics. This tunability allows practitioners to select regime-appropriate configurations to satisfy specific clinical or regulatory demands on model transparency, crucial for deployment in high-stakes medical settings.

From a theoretical perspective, the robust coverage-precision trade-off challenges the view of saliency evaluation as a monolithic objective and highlights the importance of evaluation metrics sensitive to clinical annotation granularity. Practically, the approach generalizes across disease tasks and is resilient to noisy, coarse annotations—prevalent real-world constraints.

Future research directions include extending explanation-aware objectives to multi-class/multi-label settings, leveraging more precise annotation modalities, integrating other XAI paradigms beyond Grad-CAM, and examining cross-modality robustness. The practical guidance for balancing accuracy and interpretability also points towards broader adoption in regulatory-compliant health AI systems.

Conclusion

By formally integrating explanation supervision into the model training objective, this work advances interpretability in medical imaging from post-hoc visualization to directly controllable, quantitatively evaluated model behavior. The demonstrated coverage-precision trade-off and distinct effectiveness of logit-based explanation losses provide actionable insights for XAI design under clinical supervision constraints. This framework sets a foundation for more trustworthy and clinically relevant AI, adaptable to a wide spectrum of biomedical imaging tasks and annotation regimes (2605.10054).