Quality-Guided Focal Loss: Enhancing Minority Class Detection in Haematological Imaging

In the critical race against malaria, the most dangerous parasites often hide in plain sight. When parasitaemia falls below 1%, precisely when early detection matters most, conventional AI detection systems falter despite impressive aggregate metrics. This paradox of “seeing everything except what matters most” stems from a fundamental detection dilemma: infected cells comprise a vanishingly small minority that conventional approaches systematically overlook. We propose a methodical Quality-Guided Focal Loss (QGFL), a framework that reconceptualizes how detection systems learn from imbalanced data. By integrating class-specific focusing parameters, quality-guided weighting, and spatial awareness through UIoU, QGFL achieves a remarkable improvement in detecting infected cells in the clinically vital 1–3% parasitaemia range. Our cross-dataset validation confirms QGFL’s generalizability across diverse imaging conditions without requiring dataset-specific tuning. This work advances the approach to minority class detection in medical imaging, demonstrating how prediction quality can guide model optimization, ensuring that what matters clinically also matters computationally.