3D Medical Image Segmentation Using 2D Bounding Box Localization

The patent applies segmentation by dividing the 3D medical image into multiple 2D slices along different orientations (axial, coronal, sagittal). Each slice is processed independently by 2D CNNs to generate probability maps, which are then integrated to form the final 3D segmentation. This approach improves reliability by breaking down the complex 3D problem into manageable 2D tasks while maintaining productivity through automated processing.

The patent transforms the 3D segmentation problem into a series of 2D classification problems by processing images in three orthogonal orientations. The 2D probability maps from each orientation are then integrated to reconstruct the 3D segmentation. This dimensionality change allows the use of computationally efficient 2D CNNs while achieving accurate 3D segmentation, resolving the contradiction between speed and accuracy.

The patent segments the 3D volume into multiple 2D slices and processes each slice independently using 2D CNNs. This segmentation approach reduces computational complexity compared to processing the entire 3D volume at once, thereby reducing energy consumption while maintaining segmentation precision through the integration of multiple 2D probability maps.

The patent replaces traditional image registration-based methods with a machine learning approach using 2D and 3D CNNs. This substitution eliminates the computationally intensive registration process while achieving comparable or superior segmentation precision, thereby reducing energy consumption and computational resource requirements.

The patent implements self-service by training the 2D and 3D CNN models to perform segmentation automatically without requiring manual outlining. The system learns from training data and autonomously segments new images, eliminating the need for time-consuming manual work while maintaining high accuracy through the learned features from training.

The patent performs preliminary action by training the 2D and 3D CNN models on annotated training data before deployment. This preliminary training phase enables the model to automatically perform segmentation with high accuracy, replacing the need for manual outlining during actual use and significantly improving productivity while maintaining reliability.

The patent segments the 3D volume into multiple 2D slices and processes each slice independently using computationally efficient 2D CNNs. This segmentation approach reduces the computational burden compared to direct 3D segmentation while maintaining accuracy through the integration of probability maps from multiple orientations. The method then applies a 3D CNN only to the localized bounding box region, further optimizing resource usage.

The patent transforms the computationally intensive 3D segmentation problem into a series of 2D classification problems that can be solved more efficiently. By processing images in three orthogonal orientations and integrating the results, the method achieves accurate 3D segmentation with reduced computational energy compared to direct 3D processing.

The patent processes images in three orthogonal orientations (axial, coronal, sagittal) and integrates the probability maps from each orientation to reconstruct the 3D segmentation. This multi-dimensional approach preserves 3D context information that would be lost in single-orientation 2D processing, thereby maintaining measurement precision while benefiting from the computational efficiency of 2D CNNs.

The patent merges the probability maps generated from three different orthogonal orientations to form the final 3D segmentation. By combining information from multiple views, the method preserves 3D context and improves measurement precision while maintaining the computational efficiency of 2D processing approaches.

The patent performs preliminary localization to identify a bounding box containing the target anatomy before performing detailed segmentation. This preliminary action focuses subsequent computational resources on the relevant region, reducing overall energy consumption. The localization step uses 2D CNNs on orthogonal slices to efficiently determine the bounding box coordinates.

The patent segments the processing task into two stages: (1) bounding box localization using 2D CNNs on orthogonal slices, and (2) detailed segmentation using 3D CNNs only within the localized bounding box. This segmentation of the processing pipeline reduces overall computational energy by avoiding processing of entire 3D volumes, while the added complexity is managed through modular implementation.