PneumoNet: On-Device Continual Learning for Pneumonia Diagnosis
PneumoNet, a novel method for diagnosing pneumonia at the point of care, has been created by researchers to function on medical devices in settings with limited resources. This approach integrates a compact CNN for predictions on the device, a dual-stage balanced buffer for replay that ensures class balance, and a dynamic class-weighted loss to address imbalances in training batches. When tested on a domain-shifted PneumoniaMNIST dataset that mimics five realistic scenarios of domain change, PneumoNet attained an accuracy of 86.6% with a mere 1.4% forgetting rate, outperforming existing models in size and speed. This technique mitigates performance drops due to domain shifts from variations in devices, patients, or institutions, facilitating adaptable, privacy-conscious diagnostic AI for both everyday and pandemic healthcare needs.
Key facts
- PneumoNet is a domain-incremental learning method for pneumonia diagnosis.
- It combines a lightweight CNN, dual-stage balanced buffer, and dynamic class-weighted loss.
- Evaluated on domain-shifted PneumoniaMNIST dataset with five scenarios.
- Achieves 86.6% accuracy with 1.4% forgetting.
- Smaller and faster than existing baselines.
- Designed for point-of-care devices in resource-limited settings.
- Addresses domain shifts from devices, patients, or institutions.
- Enables adaptive, privacy-preserving diagnostic AI.
Entities
Institutions
- arXiv