Geometric Embedding for Robotic Surgical Image Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image classification methods face challenges with limited and imbalanced labeled data, particularly in tasks like cancer classification, where data augmentation techniques often produce irrelevant or non-informative additional samples, leading to decreased accuracy and robustness.
Innovation Solution
The method employs geometric embedding, which generates new representations of input data using differential geometry to preserve holistic information, combined with a CNN-based deep fusion architecture that fuses class-specific activations from different modalities to produce stronger and more discriminative feature maps.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation through affine/geometric transformations is used, then the number of training samples increases, but most additional samples are not informative enough and are even irrelevant, decreasing classification accuracy
Solution Approach 1:
The patent applies parameter changes by transforming images through multiple geometric transformations (rotation, flipping, cropping, scaling) and then selecting only those transformed samples that meet specific quality criteria. This changes the parameters of the training data by filtering for informativeness rather than simply increasing quantity, thereby maintaining classification accuracy while still expanding the training set.
Solution Approach 2:
The patent applies local quality by evaluating and selecting transformed training samples based on their local informative content. Instead of treating all augmented samples uniformly, the system assesses each transformed sample's quality and selects only those that provide meaningful information for the specific classification task, ensuring that added samples contribute to accuracy rather than diluting it.
2Quantity of substance
If random patches are extracted and used as additional training samples, then more training data is available, but inter-class similarities increase when similar objects appear in different scene categories, negatively affecting accuracy and robustness
Solution Approach 1:
The patent applies parameter changes by transforming images through multiple geometric transformations (rotation, flipping, cropping, scaling) and then selecting only those transformed samples that meet specific quality criteria. This changes the parameters of the training data by filtering for informativeness rather than simply increasing quantity, thereby maintaining classification accuracy while still expanding the training set.
Solution Approach 2:
The patent applies dimensionality change by considering multiple dimensions of image transformation simultaneously (rotation angle, flip orientation, crop position, scale factor) and evaluating samples across this multi-dimensional transformation space. This allows the system to select diverse, informative samples that reduce inter-class similarity while maintaining comprehensive coverage of object variations.
3Adaptability or versatility
If data augmentation produces enormous amount of additional training data, then sample diversity increases, but additional computational resources are required to handle the data
Solution Approach 1:
The patent applies partial action by performing geometric transformations on only a subset of the training data rather than all images. The system transforms images and selects a limited number of high-quality transformed samples that meet specific informativeness criteria, avoiding the computational overhead of processing and storing enormous amounts of augmented data while still achieving diverse training coverage.
Solution Approach 2:
The patent applies discarding and recovering by generating multiple transformed versions of each training image, evaluating their quality, and discarding those that do not meet the informativeness threshold. Only the valuable transformed samples are retained for training, efficiently allocating computational resources to process and store only the most informative augmented data.
Data Source
AI summary
A method for digital image classification and localization includes receiving a digital image of a biological organism from an imaging apparatus, the digital image comprising a plurality of intensities on a 2-dimensional grid of points, generating a plurality of discriminative representations of the 2D digital image by extracting dominant characteristics of the image from three different viewpoints, where the plurality of discriminative representations form a 3-dimensional digital image, combining the 3D digital image with the 2D digital image in a convolutional neural network that outputs a 3-channel feature map that localizes image abnormalities in each of the three channels and includes a detection confidence that each abnormalities is a neoplasm, providing the 3-channel feature map to a controller of a robotic surgical device where the robotic surgical device uses the 3-channel feature map to locate the neoplasm within the biological organism in a surgical procedure for treating the neoplasm.