Accessible AI Diagnostics and Lightweight Brain Tumor Detection on Medical Edge Devices

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate brain tumor detection is critical for treatment, especially in resource-limited areas.
• Existing methods often struggle with small or variable-sized tumors and high computational demands.
• Deployment on edge devices requires lightweight and efficient AI models.

Purpose of the Study:

• To develop a lightweight and efficient RetinaNet variant for brain tumor detection on medical edge devices.
• To maintain high detection accuracy while reducing computational overhead for real-time analysis.
• To improve diagnostic accessibility in underserved regions through portable AI tools.

Main Methods:

• Replaced the ResNet backbone in RetinaNet with MobileNet.
• Utilized depthwise separable convolutions to reduce computational complexity.
• Evaluated the model on the BRATS dataset for brain tumor detection.

Main Results:

• Achieved an overall average precision (AP) of 32.1.
• Demonstrated superior performance in small tumor detection (AP_S: 14.3) and large tumor localization (AP_L: 49.7).
• Significantly reduced computational costs, enabling real-time analysis on low-power hardware with confidence scores over 81%.

Conclusions:

• The proposed lightweight RetinaNet variant offers an efficient solution for brain tumor detection on edge devices.
• The model's performance, especially in detecting small tumors, addresses key diagnostic challenges.
• This advancement can enhance early tumor detection and improve patient outcomes in remote or underserved areas.