What limits performance of weakly supervised deep learning for chest CT classification?

Models can tolerate up to 10% label error before performance declines: The study found that chest CT classification models maintained performance even when training data contained up to 10% incorrect labels, suggesting some robustness to imperfect data—a critical finding for real-world medical imaging where perfect labeling is often impossible.

More training data consistently improves performance, but with diminishing returns: Disease classification performance improved steadily as training data increased, but plateaued at around 75% of the available dataset, indicating that collecting additional data beyond a certain point may not justify the effort and cost.

Binary classification can be misleading despite better performance metrics: While binary models (single disease vs. no disease) outperformed multi-label models across all disease categories, the researchers found this superior performance was misleading—the binary models were actually learning to recognize co-occurring diseases rather than the specific disease features, raising concerns about whether models truly understand what they're classifying.

Weak supervision has real limitations that must be understood before deployment: The paper emphasizes that while weakly supervised learning with noisy labels is practical for medical imaging, practitioners need to understand these constraints and their effects on model reliability before using such systems in clinical settings.