Semantic Drift in Medical Imaging: When Model Explanations Diverge Without Accuracy Loss

Introduction

Transfer learning has become the de facto standard for medical image classification, offering reliable performance gains across radiological tasks. Yet a troubling question remains largely unexamined in clinical deployment pipelines: does the model continue to look at the same anatomical evidence after fine-tuning, or does it silently reconstruct its visual reasoning while maintaining predictive accuracy? In their recent study, "When Fine-Tuning Changes the Evidence: Architecture-Dependent Semantic Drift in Chest X-Ray Explanations," Elangovan and Ting identify a phenomenon they term semantic drift, where the attribution structure supporting model predictions undergoes systematic transformation between transfer learning and full fine-tuning phases, even when classification performance remains stable.

This finding carries significant implications for clinical AI safety. In medical contexts, explanation stability is not merely a technical curiosity but a prerequisite for trust. If a model fine-tuned for pneumonia detection shifts its attention from consolidation patterns to unrelated anatomical structures while maintaining accuracy, the clinical narrative derived from its explanations becomes unreliable. The authors demonstrate that this drift is not random but architecture dependent, and critically, that stability rankings across different neural networks can reverse depending on which attribution method one employs.

Quantifying Explanation Instability

The study establishes a rigorous methodological framework for measuring semantic drift in multi-class chest X-ray classification. Working with a five-class dataset comprising Normal, Pneumonia, Tuberculosis, COVID-19, and Lung Opacity cases (3,354 test samples), the authors evaluate three ImageNet pretrained architectures: DenseNet201, ResNet50V2, and InceptionV3. Their two-phase training protocol isolates the drift phenomenon by comparing model states at epoch 8 (transfer learning plateau with frozen backbone) against epoch 19 (fine-tuning convergence with unfrozen layers).

To ensure that measured changes reflect genuine shifts in visual reasoning rather than corrections to misclassification errors, the authors apply strict true-positive filtering. Only samples correctly classified by all three architectures at both training phases are included, yielding 2,430 samples (72.5% of the test set). This constraint is methodologically crucial; it isolates explanation evolution from predictive correctness changes, allowing the researchers to examine how models change their minds even when they remain consistently correct.

The authors propose four reference-free metrics to quantify drift. Overlap IoU (Intersection over Union) serves as the primary measure, calculating the spatial agreement between thresholded attribution maps from the transfer learning and fine-tuning phases. Spatial displacement tracks the normalized Euclidean distance between centers of mass of attribution maps. Pattern correlation computes Pearson correlation between continuous attribution values, while concentration change measures differences in Shannon entropy to detect attention redistribution. This multi-dimensional approach captures both coarse localization stability and fine-grained structural reorganization.

The Architecture-Attribution Interaction

The empirical findings reveal a nuanced picture of explanation stability. Across all architectures, coarse anatomical localization remains relatively stable; models continue to focus on lung fields rather than drifting to background regions. However, the Overlap IoU metric exposes pronounced reorganization of evidential structure within these regions, indicating that fine-tuning induces subtle but meaningful shifts in which specific visual features drive predictions.

Perhaps the most consequential finding concerns the interaction between architecture choice and attribution method. The authors evaluate explanations using both LayerCAM (which preserves fine-grained spatial detail via pixel-wise gradients) and Grad-CAM++ (which employs higher-order gradients for adaptive weighting). Strikingly, stability rankings across the three architectures reverse when switching between these methods. An architecture that appears stable under LayerCAM may demonstrate significant drift under Grad-CAM++, even after predictive performance has converged.

This observation extends prior work on explanation disagreement by demonstrating that the choice of attribution objective fundamentally alters our assessment of model stability. LayerCAM and Grad-CAM++ encode different notions of feature importance; LayerCAM emphasizes precise spatial localization through dense gradient propagation, while Grad-CAM++ aggregates gradient information to highlight influential feature map regions. The fact that stability rankings invert across methods suggests that explanation stability is not an intrinsic property of model weights, but rather an emergent property of the triplet formed by architecture, optimization phase, and attribution objective.

Beyond Accuracy: Toward Explanation-Centric Validation

The phenomenon of semantic drift raises profound questions about current validation practices in medical AI. Standard deployment pipelines optimize for and validate predictive metrics (AUC, sensitivity, specificity) while treating explanation stability as an afterthought. This study suggests that such approaches may certify models as clinically ready while overlooking fundamental shifts in their decision-making process.

Consider the clinical workflow implications. A radiologist using a fine-tuned model might establish trust based on explanations generated during the transfer learning phase, noting that the model correctly highlights consolidation patterns for pneumonia. After fine-tuning, if the model maintains accuracy but shifts attention to adjacent anatomical structures or texture patterns invisible to human perception, the radiologist continues to trust the model based on outdated assumptions about its reasoning. This creates a dangerous epistemic gap between perceived and actual model behavior.

The architecture-dependent nature of drift further complicates deployment decisions. DenseNet201, ResNet50V2, and InceptionV3 exhibit distinct connectivity patterns and feature extraction hierarchies, leading to differential susceptibility to semantic drift during gradient updates. This suggests that architecture selection criteria for medical imaging should incorporate explanation stability metrics alongside traditional performance benchmarks. A model with marginally lower AUC but higher explanation stability may be preferable for clinical deployment where interpretability requirements are stringent.

Limitations and Open Questions

While the study provides a robust framework for measuring drift, several limitations warrant consideration. The analysis focuses exclusively on gradient-based attribution methods, leaving unclear whether similar drift patterns emerge under perturbation-based or concept-based explanation techniques. Additionally, the study evaluates drift purely through computational metrics without human-grounded validation; we do not know whether drifted explanations remain clinically valid or represent genuine deterioration of visual reasoning.

The inverse-frequency weighting scheme, while necessary for handling class imbalance (with Tuberculosis representing only 4.2% of samples versus Lung Opacity at 35.8%), aggregates drift metrics across pathologies with vastly different visual manifestations. Future work might investigate whether semantic drift occurs uniformly across pathologies or concentrates in specific disease categories with subtle radiographic features.

Open questions remain regarding the functional significance of drift. Does semantic drift correlate with robustness to domain shift, or does it indicate overfitting to training set idiosyncrasies? Can we regularize fine-tuning procedures to minimize drift while maintaining accuracy gains? The authors' formalization of drift as a cross-phase evidence transformation offers a mathematical foundation for developing such regularization techniques, potentially enabling optimization objectives that penalize attribution map reorganization beyond certain thresholds.

Conclusion

"When Fine-Tuning Changes the Evidence" establishes that explanation stability cannot be assumed from predictive performance alone. The study demonstrates that fine-tuning, despite its benefits for task-specific accuracy, induces architecture-dependent semantic drift in the visual evidence supporting chest X-ray classifications. The observation that stability rankings reverse across attribution methods underscores the need for multi-method explanation consensus in clinical validation pipelines.

As medical AI systems transition from research environments to clinical practice, we must expand our validation frameworks to treat explanation stability as a first-class metric. The triplet of architecture, optimization phase, and attribution method determines not just what a model predicts, but how it reasons. Understanding and controlling semantic drift represents an essential step toward deploying AI systems whose decision-making processes remain transparent and consistent throughout their operational lifecycle. Future research should explore human-centered validation of drifted explanations and develop training protocols that explicitly optimize for both predictive accuracy and explanation stability.