A Two-Tier Adaptive Preprocessing Framework for Mitigating Extreme Noise and Class Imbalance in Buccal Mucosa Comet Assay Images
Abstrak
Comet assay-based genotoxicity assessment in buccal mucosa cells shows strong clinical potential. However, computational analysis is hindered by visual complexity from cellular debris and significant class imbalance. In these noisy settings, single-pass classification models often face the Accuracy Paradox, showing superficially high performance by mainly identifying the majority class while missing severe DNA damage (Class 4 comets). To address this, we propose a two-tier classification framework that strategically separates upstream noise filtration from downstream multi-class grading. We systematically evaluate architectural trade-offs between machine learning-based filters—specifically Logistic Regression (LR) and Support Vector Machine (SVM), optimized for feature-space purity—and a broad Descriptive Statistics (DS) boundary, which prioritizes the retention of degraded biological structures. Integrated with cost-sensitive optimization and SMOTE-Tomek balancing—which serves as a robust compromise for stabilizing minority variance—our proposed framework was rigorously benchmarked. By coupling an upstream LR filter with the Tier-2 SVM detector, the binary module outperforms OpenComet and the previous GamaComet, achieving a held-out test F1-Score of 0.790 and an AUROC of 0.931. In the downstream multi-class phase, pairing the broader DS filter with the SMOTE-Tomek-balanced SVM classifier yielded highly significant statistical improvements (<inline-formula> <tex-math notation="LaTeX">$p \lt 0.0001$ </tex-math></inline-formula>) over historical baselines, reaching a Macro F1-Score of 0.557 and a Macro AUROC of 0.916. This highlights the need for upstream leniency to provide the downstream classifier with a wider representation of artifacts, maximizing the potential to retain critically damaged Class 4 comets. While currently bounded by a single-institution dataset, our findings establish a configurable diagnostic gatekeeper that overcomes vulnerabilities of previous tools and sets robust preprocessing standards for future multi-institutional Deep Learning integration.
Identifier
- DOI: https://doi.org/10.1109/access.2026.3714506
- OpenAlex: https://openalex.org/W7169615242