Neural Network Training for Medical Image Lesion Detection

The neural network processes images at multiple resolutions by segmenting the input into different scale representations. The system divides the image processing task into multiple pathways handling different resolution levels, allowing comprehensive feature extraction without requiring the entire network to process all high-resolution details simultaneously, thus managing computational complexity while improving image quality assessment.

The patent introduces a multi-scale dimension to the image processing by incorporating images at different resolutions (first resolution and second resolution). This dimensional approach allows the network to capture both global and local features, improving measurement precision through multi-resolution analysis while distributing computational load across different scale levels rather than concentrating it at a single high-resolution level.

The patent applies preliminary data-augmenting operations including cropping, rotating, and flipping to the training images before neural network training. These preprocessing actions are performed in advance during the training phase, allowing the network to learn from pre-enhanced data that already emphasizes pathological features. This preliminary action reduces the need for extensive real-time processing during actual diagnosis, thus improving diagnostic accuracy while managing processing time.

The system applies multiple data-augmenting operations (cropping, rotating, flipping, scaling) that may seem excessive but are strategically used during training to over-emphasize pathological features. During actual inference, the network has already learned to recognize these features from the augmented training data, so minimal processing is needed at runtime. This partial or excessive action during training improves reliability without proportionally increasing processing time during deployment.

The training process segments images into multiple scale versions (first resolution and second resolution) and processes them through separate but coordinated neural network pathways. Each pathway is trained on specific resolution levels, allowing the model to learn scale-specific features independently. This segmentation of the training task reduces overall training complexity by breaking down the multi-scale problem into manageable sub-tasks while maintaining high lesion detection accuracy through the integration of multi-scale features.

The patent employs a universal neural network architecture that can process images at multiple resolutions through the same fundamental structure. The network is designed with multi-functionality to handle both first-resolution and second-resolution images using shared computational components. This universal design reduces model training complexity by avoiding the need for entirely separate models for different scales, while still achieving high lesion detection accuracy through multi-scale feature extraction.