System and method of processing medical images
Inventors
Jackson, Brian H. • Stavish, Coleman C. • Jing, Yating • Kulp, John
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-10346980-B2
Publication Date
2019-07-09
Expiration Date
Abstract
Presented are techniques for processing medical images. The techniques can include accessing a stored medical image and electronically representing a plurality of overlapping tiles that cover the medical image, each overlapping tile including a non-overlapping inner portion and an overlapping marginal portion. The techniques can also include in parallel, and individually for each of a plurality of the overlapping tiles: applying a segmentation process to identify objects in the at least one medical image, identifying inner object data representing at least one inner object that is contained within an inner portion of at least one tile, and identifying marginal object data representing at least one marginal object that overlaps a marginal portion of at least one tile. The techniques can also include merging at least some of the marginal object data to produce merged data, and outputting object data including the inner object data and the merged data.
Core Innovation
The problem is how to process at least one electronically stored medical image in a scalable manner while dealing with artifacts and object segmentation across image boundaries. The approach represents the medical image as a plurality of overlapping tiles, where each overlapping tile comprises a non-overlapping inner portion and an overlapping marginal portion, enabling parallel handling of tile-local content. Objects are identified in each tile using a segmentation process, and the method separates inner object data contained within inner portions from marginal object data overlapping marginal portions.
For the marginal regions, the method merges at least some of the marginal object data to produce merged data, so that final output includes the inner object data and the merged data. In addition, marginal object merging addresses overlap redundancy between tiles using procedures such as splitting overlapping marginal object data, stitching the split data, smoothing the stitched data, and performing recursive merging for marginal object data based on overlapping tile participation. This supports producing a consolidated representation of objects that cross tile boundaries.
The processing pipeline further enables downstream visualization and scalable access by building a spatial access tree that indexes dense compressed object chunks and supports bounding-box queries from client computers over a computer network. The system retrieves only needed object data from persistent storage based on identifying data returned from the clients. Post-processing can also decluster improperly clustered objects by computing curvature data, identifying clustered objects from the curvature data, and applying an adaptive watershed technique to split clustered objects into at least two unclustered objects.
Claims Coverage
The independent claims cover a computer-implemented method and a system that implement tiled, parallel segmentation with inner/marginal separation and marginal-object merging, producing output object data containing inner objects and merged marginal results. The independent claim set includes inventive features that distinguish inner from marginal tile portions, merge marginal object data for overlaps, and optionally incorporate declustering using curvature data with adaptive watershed splitting and spatial access-tree querying over a network.
Overlapping tile representation with inner and marginal portions
Electronically representing a plurality of overlapping tiles that cover the at least one medical image, each overlapping tile comprising a non-overlapping inner portion and an overlapping marginal portion.
Parallel per-tile segmentation with inner and marginal object data
In parallel, and individually for each of a plurality of the overlapping tiles, applying a segmentation process to identify objects in the at least one medical image, identifying inner object data representing at least one inner object contained within an inner portion of at least one tile, and identifying marginal object data representing at least one marginal object that overlaps a marginal portion of at least one tile.
Merging marginal object data and outputting inner plus merged results
Merging at least some of the marginal object data to produce merged data, and outputting object data comprising the inner object data and the merged data.
System with persistent memory and processors for tiled segmentation and merging
A system comprising an electronic persistent memory and at least one electronic processor, where the persistent memory stores at least one medical image and the processor is configured to represent the image as overlapping tiles with non-overlapping inner portions and overlapping marginal portions, perform parallel per-tile segmentation with inner and marginal object data, merge at least some marginal object data to produce merged data, and output object data comprising the inner object data and the merged data.
Declustering using curvature data and adaptive watershed splitting
Computing curvature data for a clustered object, determining from the curvature data that the object is a clustered object, and applying an adaptive watershed technique to split the clustered object into at least two unclustered objects.
Spatial access tree and bounding-box based network retrieval
Removing object data from a spatial access tree structure, storing the object data in persistent memory, providing the spatial access tree over a computer network to client computers that query it using bounding box data to obtain identifying data for included objects, receiving identifying-based query data back from the clients, and providing the corresponding object data for the identified objects within the bounding box represented by the bounding box data.
Across the independent claims, the central coverage is a tiled, parallel processing scheme that distinguishes inner from marginal tile regions, segments objects per tile, merges marginal object data from overlaps into merged data, and outputs combined object data including inner and merged results. Additional inventive narrowing in dependent claim coverage includes declustering based on curvature data with an adaptive watershed technique and spatial access-tree based bounding-box querying in a client-server setting.
Stated Advantages
Parallelization to enable scalable processing of medical images.
Reduced computation/memory load.
Improved accuracy via local heterogeneity handling.
>30x speedup.
Documented Applications
Processing and analyzing large medical images, including whole-slide images, using overlapping tiling with object segmentation and merged object output.
Declustering improperly clustered objects by curvature-based detection and adaptive watershed splitting.
Client-server viewing and retrieval of selected objects over a computer network using bounding-box queries indexed by a spatial access tree.
Interested in licensing this patent?