Adaptation Depth Over Synthetic Augmentation for Classification of Imbalanced Circulating Tumor Cells

Abstract

Circulating tumor cells (CTCs) are a non-invasive biomarker of metastasis. However, they are extremely rare, making automated detection in fluorescence microscopy images a severely class-imbalanced learning problem with a small, fixed-size positive class. A common approach is to synthesize additional minority examples with a generative model. On the hardest subset of a CTC dataset (positive class = abnormal CTCs, 106 patches vs. a negative class restricted to the single most CTC-like “suspicious” subtype; imbalance $\approx 5: 1)$, we investigate whether diffusion-based synthetic augmentation is effective. Each configuration is evaluated over 10 data splits (5-fold ×2 repeats). Across synthetic dose, generation guidance, and a transfer-based generator, synthetic augmentation does not significantly outperform simple class balancing on the threshold-independent primary metric: the area under the precision-recall curve (PR-AUC). PR-AUC is instead governed by how deeply the ImageNet-pretrained backbone is adapted: across splits, it rises monotonically with unfreeze depth, from 0.887 (frozen) to 0.965 (full). The frozen head performs significantly worse, making adaptation depth the only intervention with a statistically robust across-split lift. The same frozen-versus-adapted axis governs the generator: a pretrained diffusion model fine-tuned deeply enough passes a train-on-synthetic/test-on-real (TSTR) fidelity gate, yet still does not improve downstream classification, so passing TSTR does not imply a downstream gain. Imbalance losses and threshold tuning shift the operating point without changing the curve. On this task and modality, performance is bound by representation capacity.