Latent Space Analysis for Interpretable Uncertainty in Melanoma Classification

Technical Summary: Latent Space Analysis for Interpretable Uncertainty in Melanoma Classification

Problem Statement
Melanoma is a highly aggressive skin cancer where early detection is critical for patient survival. While deep learning has advanced skin lesion classification, standard models often operate as unexplainable "black boxes." These models struggle to transparently communicate diagnostic uncertainty, particularly in borderline cases where lesions share morphological features of both nevi (NV) and melanomas (MEL). This lack of interpretability limits clinical trust and hinders the effective integration of automated tools into diagnostic workflows, where understanding the "why" behind a prediction is as important as the prediction itself.

Methodology
The authors propose a hybrid framework that transcends simple binary classification by leveraging a structured generative latent space for interpretable decision support. The methodology consists of three primary components:

1. Class-Aware Adversarial Variational Autoencoder (VAE-GAN):

   • Architecture: The model employs an encoder-decoder structure. The encoder maps input images (128×128×3) to a continuous 256-dimensional latent space, outputting a mean vector ($\mu$) and log-variance ($\log(\sigma^2)$). The reparameterization trick ($z = \mu + \sigma \odot \epsilon$) is used to ensure differentiability.
   • Adversarial Training: To overcome the tendency of standard VAEs to produce blurry reconstructions, a discriminator is integrated. This adversarial component forces the decoder to preserve high-frequency morphological details critical for dermatological assessment.
   • Class Awareness: An auxiliary Multi-Layer Perceptron (MLP) is attached to the latent mean ($\mu$) to enforce class-discriminative clustering (MEL vs. NV) during training.
   • Loss Function: The model is optimized via a multi-objective loss function: $L_{total} = L_{recon} + \lambda_{KL}L_{KL} + \lambda_{clf}L_{clf} + \lambda_{adv}L_{adv}$. This balances reconstruction quality, latent space regularization, classification accuracy, and adversarial realism.

2. Latent Feature Extraction and Classification:

   • Post-training, the encoder serves as a deterministic feature extractor, outputting the latent mean vector $\mu$ (omitting the stochastic noise term) for each image.
   • These 256-dimensional vectors are standardized and fed into three classifiers: Support Vector Machine (SVM), Random Forest (RF), and XGBoost.
   • Hyperparameters were tuned via Grid Search with 3-fold cross-validation, optimizing for macro-averaged F1-score. XGBoost was selected as the final classifier.

3. Interpretable Uncertainty via Content-Based Image Retrieval (CBIR):

   • Instead of providing only a binary label, the XGBoost classifier outputs a continuous probability score ($p \in [0, 1]$).
   • For cases with high uncertainty (scores near 50%), the framework triggers a CBIR system. The query image is projected into the latent space, and the $k$ nearest neighbors (based on Euclidean distance) are retrieved from the training archive.
   • This allows clinicians to visually compare an ambiguous lesion against biopsy-confirmed historical precedents, providing evidence-based context for the uncertainty.

Key Contributions

• Paradigm Shift: The work repurposes the generative formulation of VAEs not for image synthesis, but to distill clinical features into a continuous, semantically structured latent space.
• Interpretable Uncertainty: The framework explicitly translates algorithmic hesitation into a visual decision support tool. By retrieving morphologically similar historical cases for borderline predictions, it bridges the gap between predictive performance and clinical trust.
• Hybrid Architecture: The integration of an adversarial VAE with an auxiliary classifier and a gradient-boosted tree (XGBoost) creates a system that is robust to class imbalance while maintaining a structured topology for retrieval.

Results
The framework was evaluated on the ISIC 2019 dataset, focusing on the binary classification of Nevi (12,875 training, 2,495 test) and Melanomas (4,522 training, 1,327 test).

• Predictive Performance: The VAE+XGBoost model achieved an accuracy of 81.26%, a macro-averaged precision of 0.79, recall of 0.80, and an F1-score of 0.80. It attained an AUC of 0.8689.
• Comparison to Baselines: While end-to-end convolutional models like ResNet50 and DenseNet121 achieved slightly higher AUCs (~0.90) and F1-scores (0.83), the proposed framework performed competitively, outperforming ConvNeXt-T (AUC 0.66) and closely trailing EfficientNetV2 (AUC 0.88). The authors note that the slight reduction in raw predictive metrics is a deliberate trade-off for interpretability.
• Retrieval Efficacy: The CBIR system demonstrated strong semantic consistency. For general queries, it achieved a Mean Precision@5 (mP@5) of 72.0% and Top-5 Accuracy of 93.9%. Crucially, for true melanoma queries (the minority class), the system achieved an mP@5 of 44.7%, significantly outperforming random retrieval and providing a validated "red flag" mechanism for clinicians.

Significance and Claims
The paper claims that its primary significance lies in moving beyond the "black box" paradigm. By utilizing a continuous latent space, the system does not merely classify lesions but provides a mechanism to visualize why a classification is uncertain. The retrieval of biopsy-confirmed precedents for borderline cases offers a dual benefit: it aids the clinician in finalizing a diagnosis through visual comparison and acts as an early warning system for lesions sharing features of both nevi and melanomas. The authors conclude that this approach enhances human diagnostic confidence by keeping the clinician "in the loop," offering transparent, evidence-based visual support rather than forcing a blind binary prediction. Future work is planned to expand the framework to multi-class diagnosis (including BCC and BKL) and to explore advanced latent space variants like $\beta$-VAE for improved feature disentanglement.