Addressing Class Imbalance Issues in Haematological Images using Repeat Factor Sampling

Applying deep learning to medical imaging, especially in haematology, faces significant challenges due to class imbalance, where infected cells are vastly outnumbered by normal cells. This study addresses this issue using a Customized Repeat Factor Sampling (CRFS) method integrated into the Faster R-CNN architecture within the Detectron2 framework, with malaria detection as a use case. By dynamically adjusting sampling weights based on the number of infected instances, CRFS significantly improves model performance. Results show notable increases in precision, recall, and F1 scores for detecting malaria-infected cells, demonstrating the method's effectiveness in enhancing detection accuracy. This approach offers a straightforward and computationally efficient solution to class imbalance, with potential applications across various haematological disorders, improving screening processes for other rare blood conditions.