A Study on Energy Consumption in AI-Driven Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Sustainable Computing

Background:

• The environmental impact of artificial intelligence (AI) in medical image analysis is under-explored.
• AI models, particularly for segmentation tasks, require significant computational resources.
• Understanding energy consumption is crucial for sustainable AI development in healthcare.

Purpose of the Study:

• To analyze the energy demands of AI workflows for medical image segmentation.
• To compare energy consumption between AI training and inference phases.
• To evaluate the impact of different convolutional variants and optimization techniques on energy efficiency.

Main Methods:

• Utilized the Kidney Tumor Segmentation-2019 (KiTS-19) dataset for analysis.
• Evaluated Standard Convolution, Depthwise Convolution, and Group Convolution.
• Assessed optimization techniques including Mixed Precision and Gradient Accumulation.
• Measured energy consumption focusing on computational complexity, memory access, and I/O operations.

Main Results:

• AI training is energy-intensive, but inference's cumulative energy use can be higher over a model's lifecycle.
• Depthwise Convolution combined with Mixed Precision demonstrated the lowest training energy consumption with strong performance.
• Group Convolution exhibited poor energy efficiency due to substantial input/output overhead.
• Inference energy consumption is significantly influenced by recurring computational demands.

Conclusions:

• Depthwise Convolution with Mixed Precision presents the most energy-efficient configuration for medical image segmentation AI.
• GPU-centric strategies and energy-conscious AI practices are essential for sustainable medical imaging AI.
• Actionable guidance is provided for developing scalable and environmentally responsible AI in medical image analysis.