MalariaNet: A Microcontroller-Deployable Malaria-Microscopy Detector for Point-of-Care Biosensing Under Leakage-Free...

Area of Science:

• Medical Diagnostics
• Computer Science
• Parasitology

Background:

• Compact malaria detectors are crucial for point-of-care diagnostics.
• Current benchmarking often uses the NIH Malaria dataset with per-cell random splits.
• This method can leak slide identity, potentially inflating performance metrics.

Purpose of the Study:

• To investigate the impact of data leakage in benchmarking malaria detection models.
• To propose and validate a slide-disjoint evaluation protocol.
• To introduce MalariaNet, a robust, microcontroller-deployable malaria detector.

Main Methods:

• A slide-disjoint evaluation protocol was implemented to prevent data leakage.
• Eight different microcontroller-deployable malaria detection architectures were tested.
• MalariaNet, a compact deep learning model, was developed and evaluated on-chip.

Main Results:

• Per-cell random splits significantly overestimate model performance compared to slide-disjoint evaluation.
• Headline accuracy dropped from 97.1% to 95.6% under the leakage-free protocol.
• MalariaNet achieved 95.6% accuracy with a 23.5 KB model size, running at 1.2 FPS on an STM32H743 microcontroller.

Conclusions:

• Slide-disjoint evaluation is essential for accurate benchmarking of malaria detection models.
• MalariaNet demonstrates high accuracy and interference robustness for on-device malaria detection.
• The findings necessitate a standardized, leakage-free evaluation approach for future research.