RBCNet

RBCNet employs a dual deep learning architecture to detect and count red blood cells in thin blood smear microscopy images for malaria diagnostics.

Key Features:

• Dual Deep Learning Architecture: A two-stage pipeline uses U-Net for cell-cluster or superpixel segmentation followed by Faster R-CNN to refine detections within segmented clusters.
• U-Net Segmentation: U-Net produces foreground cell-cluster masks that handle cell fragmentation and small-object or fine-scale morphological structures.
• Faster R-CNN Refinement: Faster R-CNN focuses on small cell objects within segmented clusters to enhance detection precision and reduce false positives.
• Cell Clustering Segmentation: Segmentation is performed via cell clustering rather than region proposals, enabling scalability and adaptability to varying image scales and very large images.
• Training with Non-overlapping Tiles: The model can be trained using non-overlapping tiles to optimize computational resource usage during training.
• Inference Scale Adaptation and Low Memory Footprint: During inference the model adapts to the scale of cell clusters while maintaining a low memory footprint for resource-constrained environments.
• High Accuracy: The approach achieves detection accuracy exceeding 97% with improved true positive rates and reduced false alarms compared to traditional methods.

Scientific Applications:

• Malaria diagnostics: Automated detection and counting of red blood cells in thin blood smears to support identification of infected cells and diagnostic screening.
• Large-scale clinical image analysis: Processing extensive datasets, as demonstrated on nearly 200,000 labeled cells across 965 images from 193 patients in Bangladesh, for scalable quantitative RBC analysis in research and clinical studies.

Methodology:

Initial segmentation of cell clusters using U-Net to generate foreground masks; refinement of detections within those masks using Faster R-CNN; training with non-overlapping tiles; inference that adapts to cell-cluster scale with a low memory footprint; segmentation implemented via cell clustering rather than region proposals.