Towards Robust Cancer Diagnosis: A Hybrid Deep Learning Pipeline for Imbalanced Multi-Label Microscopic Image Analysis

This paper presents a robust and flexible hybrid pipeline for multi-label classification of cell images in the context of imbalanced biomedical datasets. Using ConvNeXt, our approach integrates rare-aware techniques: conditional CutMix for under-represented labels, basic augmentations, and threshold tuning to enhance predictive performance. The pipeline further incorporates Strictly Proper Asymmetric Loss (SPA) with a Label Pairwise Regularizer (LPR), enhancing the model’s ability to produce calibrated probability estimates and consistent predictions for correlated labels. Extensive ablation studies demonstrate the framework’s consistently competitive results across various configurations. These findings highlight the effectiveness of the pipeline in tackling class imbalance and complex decision boundaries for automated biomedical image analysis.