Knockout Autoencoder for Biomedical Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current mechanisms for biomedical image segmentation, particularly using convolutional neural networks (CNNs), face challenges in accurately identifying irregular shapes and anomalies such as tumor masses and lesions, which vary in geometry and intensity, as they do not effectively segment non-regular structures.
Innovation Solution
The implementation of a knockout autoencoder engine that repurposes a U-net to predict original image contents by training on images with random noise patches, allowing the network to learn and highlight abnormal areas by computing probability distributions for each pixel based on the rest of the image, thereby detecting anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional CNNs are used for biomedical image segmentation, then regular structures can be segmented, but irregular shapes and anomalies cannot be accurately identified
Solution Approach 1:
Instead of training the network to directly segment anomalies (which are rare and irregular), the patent inverts the approach by training the network to reconstruct normal tissue structures. The autoencoder learns what normal tissue should look like, and anomalies are detected as deviations from this learned normal pattern. This inversion allows the system to handle irregular shapes effectively by focusing on the common case of normal tissue rather than the rare case of anomalies.
Solution Approach 2:
The patent changes the fundamental parameter being optimized from direct anomaly segmentation to reconstruction error minimization. By training the network to minimize reconstruction error for normal tissues while allowing high error for anomalies, the system adapts to detect irregular structures. The knockout mechanism further changes parameters by randomly masking regions to force the network to learn contextual relationships and structural patterns.
2Manufacturing precision
If U-net architecture is used for segmentation, then precise segmentations can be achieved with fewer training images, but the network cannot effectively identify irregular shapes and anomalies
Solution Approach 1:
The patent makes the U-net architecture multi-functional by using it both for image reconstruction (autoencoder function) and anomaly detection. The same network that learns to reconstruct normal tissue structures is then used to identify anomalies through reconstruction error. This universal application allows the network to leverage its segmentation capabilities while simultaneously detecting irregular shapes and anomalies that conventional single-function networks miss.
Solution Approach 2:
The reconstructed image serves as an intermediary between the input image and anomaly detection. Instead of directly detecting anomalies from the original image, the system first generates a reconstructed version representing normal tissue, then compares the original and reconstructed images to identify deviations. This intermediary reconstruction process enables reliable anomaly detection while maintaining the U-net's precision advantages.
3Loss of information
If the network is trained to predict original image from input image, then probability distributions for each pixel can be computed, but the system complexity increases
Solution Approach 1:
The patent extracts the essential function of anomaly detection from the complex task of direct segmentation. By taking out the reconstruction task and training the network separately for this purpose, the system simplifies the overall approach. The network focuses on learning normal tissue patterns through reconstruction rather than simultaneously learning to segment diverse anomaly types, reducing the effective complexity while preserving detection capability.
Solution Approach 2:
The autoencoder creates a copy (reconstruction) of the input image that represents what normal tissue should look like. This copied reconstructed image serves as a reference for detecting anomalies without requiring the network to directly model complex anomaly structures. The copying mechanism allows the system to handle information about expected pixel values while managing complexity by working with simplified reconstruction tasks rather than direct anomaly classification.
Data Source
AI summary
A mechanism is provided in a data processing system comprising a processor and a memory, the memory comprising instructions that are executed by the processor to specifically configure the processor to implement a knockout autoencoder engine for detecting anomalies in biomedical images. The mechanism trains a neural network to be used as a knockout autoencoder that predicts an original based on an input image. The knockout autoencoder engine provides a biomedical image as the input image to the neural network. The neural network outputs a probability distribution for each pixel in the biomedical image. Each probability distribution represents a predicted probability distribution of expected pixel values for a given pixel in the biomedical image. An anomaly detection component executing within the knockout autoencoder engine determines a probability that each pixel has an expected value based on the probability distributions to form a plurality of expected pixel probabilities. The anomaly detection component detects an anomaly in the biomedical image based on the plurality of expected pixel probabilities. An anomaly marking component executing within the knockout autoencoder engine marks the detected anomaly in the biomedical image to form a marked biomedical image and outputs the marked biomedical image.