LIME

LIME treats the original model as a complete black box. It does not require access to internal gradients, weights, or architecture details. This makes it universally applicable to any classifier or regressor, from deep neural networks and gradient-boosted trees to proprietary APIs. The only requirement is the ability to probe the model with perturbed inputs and observe the corresponding outputs, making it ideal for auditing third-party diagnostic systems.

LIME generates explanations by approximating the complex decision boundary locally around a single prediction. It creates a neighborhood of perturbed samples—such as randomly occluded superpixels in a medical image—and weights them by their proximity to the original instance. An inherently interpretable model, typically a sparse linear model or a decision tree, is then trained on this local dataset. The coefficients of this simple model directly quantify each feature's contribution to that specific prediction.

For medical imaging tasks, LIME does not explain individual pixels but rather contiguous superpixels—homogeneous patches of the image grouped by color similarity and spatial proximity. This is crucial for radiological interpretability because it generates explanations that align with anatomical structures rather than producing noisy, pixel-level saliency maps. A radiologist reviewing a chest X-ray explanation sees highlighted anatomical regions (e.g., a lung lobe or a mass margin) rather than scattered individual pixels.

The final explanation is a sparse linear combination of the most influential superpixels. LIME enforces sparsity by selecting a small number of features (e.g., the top 5 superpixels) that best approximate the black-box prediction. For a diagnostic model classifying a skin lesion as malignant, the explanation might state: 'Prediction = 0.92 (malignant) because of region A (+0.45), region B (+0.30), and region C (+0.15).' This additive feature attribution format is directly auditable by clinicians and regulatory reviewers.

A known limitation of LIME is the stability-plasticity trade-off in its sampling process. Because explanations are generated from random perturbations, running LIME twice on the same instance can produce slightly different superpixel weights. This non-determinism poses challenges for regulatory audit trails in medical software. Mitigation strategies include fixing the random seed, increasing the number of perturbed samples, or using deterministic variants. Evaluating LIME explanations with a faithfulness score is essential before clinical deployment.

Beyond medical imaging, LIME extends natively to tabular clinical data and radiology report text. For tabular data, perturbations involve toggling binary features or sampling from quantile-based distributions. For text, LIME removes individual words or tokens to measure their impact on the prediction. This multi-modal capability allows a unified explainability framework across a diagnostic pipeline—explaining an image classification, a structured EHR-based risk score, and a generated report summary using the same core algorithm.

A quantitative metric that evaluates how accurately an explanation reflects the model's true reasoning. It measures the correlation between attributed importance and the actual change in model output when features are perturbed. Key evaluation approaches include:

• Deletion metrics: Removing top-attributed pixels should cause a sharp drop in confidence
• Insertion metrics: Adding top-attributed pixels to a blank canvas should rapidly restore prediction
• Comprehensiveness: The difference in output when important features are removed

LIME's fidelity is often benchmarked using these scores.

An explanation that answers 'what if' questions by identifying the minimal change to an input that would alter the prediction. In diagnostic imaging, a counterfactual might show that slightly enlarging a lesion's border or changing its texture would flip a benign classification to malignant. This complements LIME's feature importance approach by providing actionable, contrastive insights that clinicians can intuitively understand—'this scan would be classified as malignant if this region were denser.'

A human-AI collaboration paradigm where medical professionals actively review AI-generated explanations before making final diagnoses. LIME's superpixel-based explanations are designed for this workflow—radiologists can visually verify whether highlighted regions correspond to clinically relevant anatomy rather than artifacts. This approach supports trust calibration, preventing both over-reliance on plausible-looking but unfaithful saliency maps and under-reliance on accurate AI assistance.