Medical Image ROI Detection Using Superpixel Classification and Extended Hough Transform
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical image processing techniques struggle to accurately detect regions of interest in digital medical images, particularly due to the large visual variability in collimated images, which limits the accuracy of collimation detection and often results in false positives when medical implants are present, and are constrained to rectangular regions of interest, failing to handle non-rectangular collimation effectively.
Innovation Solution
The method employs superpixelation using Simple Linear Iterative Clustering (SLIC) to over-segment images, followed by a machine learning-based approach to classify superpixel pairs into three classes, generating an edge map, and applying an extended Hough transform to derive the border of the region of interest, excluding false positives and identifying the optimal quadrilateral for cropping the image.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If unsupervised models for edge detection are used, then the detection process is simple, but the accuracy is limited and false positives occur when medical implants are present
Solution Approach 1:
The patent introduces superpixels as an intermediary representation between raw image data and edge detection. By first segmenting the image into superpixels and then analyzing superpixel pairs, the method creates an intermediate layer that filters out noise from medical implants while preserving true collimation boundaries, thereby improving accuracy without proportionally increasing complexity
Solution Approach 2:
The patent segments the image into multiple superpixels before performing edge detection. This segmentation approach divides the complex detection task into smaller, more manageable units (superpixel pairs), allowing for more precise local analysis and reducing false positives from implants while maintaining computational feasibility
2Device complexity
If rectangular ROI constraints are applied, then the detection algorithm is simpler, but non-rectangular collimation cannot be handled effectively
Solution Approach 1:
The patent employs a dynamic programming approach to find the optimal quadrilateral ROI, allowing the solution to adapt to different collimation shapes. The algorithm dynamically selects the best quadrilateral configuration based on the detected edges and superpixel probabilities, enabling it to handle both rectangular and non-rectangular collimation effectively without requiring separate algorithms for each shape
3Extent of automation
If two-class pixel-level classification is used, then the method can detect regions, but the accuracy is limited due to large visual variability and feature distribution overlap
Solution Approach 1:
The patent transitions from pixel-level classification to superpixel-pair-level classification, adding a new dimension to the analysis. By considering pairs of neighboring superpixels and their relationships, the method captures more contextual information and reduces the impact of visual variability and feature overlap that plague single-pixel approaches
Solution Approach 2:
The patent applies different classification strategies to different types of superpixel pairs (e.g., pairs likely to be inside ROI versus pairs likely to be outside). By tailoring the analysis to local characteristics of each superpixel pair, the method improves accuracy by focusing computational resources on discriminative regions while maintaining automated detection
Data Source
AI summary
A method for automatically detecting a region of interest in a digital medical image, comprising over-segmenting the image into a plurality of superpixels through use of an over-segmentation algorithm; for each pair of neighboring superpixels in the plurality of superpixels, computing, through a machine learning algorithm, the probability of each pair being in one of three predetermined classes; for each superpixel in the plurality of superpixels, computing a probability of the superpixel being in the region of interest; generating an edge map from computing each pixel's value based on the computed superpixel probabilities; applying an extended Hough transform to the generated edge map to generate a Hough parameter counting space; determining the optimal quadrilateral in the Hough parameter counting space by excluding false positive edges; and designating the region of interest as being within the boundary of the determined optimal quadrilateral.