Knowledge Distillation for Low-Noise Medical Image Enhancement

The patent introduces an intermediary processing step between image acquisition and final output. A trained machine learning model acts as a mediator that receives low-quality images with noise and artifacts, processes them through learned transformations, and outputs enhanced images with improved SNR and CNR. This intermediary model bridges the gap between current imaging capabilities and desired image quality.

The patent transforms the image data by changing key parameters such as signal-to-noise ratio and contrast-to-noise ratio through machine learning-based processing. The model learns optimal parameter transformations during training on paired datasets, enabling it to adjust noise levels, contrast enhancement, and overall image quality metrics without requiring changes to the physical imaging equipment.

The patent performs preliminary action by pre-training machine learning models on extensive datasets of paired low-quality and high-quality images before deployment. This offline training phase allows the model to learn complex noise patterns and enhancement strategies in advance, so that during real-time surgical procedures, the model can rapidly process images without requiring lengthy processing times or sacrificing quality.

The patent creates a virtual copy of the ideal high-quality imaging process through the trained machine learning model. Instead of physically acquiring images at higher quality settings (which would increase scan time), the model generates a computational copy of what a high-quality image would look like, based on the rapidly acquired low-quality input. This copying approach maintains productivity while achieving enhanced image quality.

The patent performs the computationally intensive work of learning optimal noise reduction strategies during an offline training phase. The machine learning model is trained on large datasets using complex algorithms and significant computational resources beforehand. Once trained, the model contains compressed knowledge that can be rapidly applied during real-time processing, transferring the computational burden from the deployment phase to the training phase.

The patent replaces traditional mechanical or algorithmic image processing methods with a data-driven machine learning approach. Instead of applying fixed mathematical filters or iterative optimization algorithms during processing, the system uses a pre-trained neural network that has learned effective processing patterns from data. This substitution enables faster processing while maintaining or improving noise reduction performance.