200 likes | 331 Views
Region of Interest Queries in CT Scans Matthias Schubert 1 Joint work with Alexander Cavallaro 2 , Franz Graf 1 , Hans-Peter Kriegel 1 , Marisa Thoma 1 1 Ludwig- Maximilians - Universität München , Database Group 2 Imaging Science Institute, University Hospital Erlangen. Outline.
E N D
Region of Interest Queries in CT Scans Matthias Schubert1 Joint work with Alexander Cavallaro2, Franz Graf1, Hans-Peter Kriegel1, Marisa Thoma1 1 Ludwig-Maximilians-UniversitätMünchen, Database Group 2 Imaging Science Institute, University Hospital Erlangen
Outline • ROI Queries on CT-Scans • ROI Retrieval Based On a General Height Scale: • Simple Solution based on Similarity Search • Solution based on Generalized Height Scale • kNN Regression for mapping slices • Iterative interpolation • Experimental Validation • Summary
Computer Tomography:CT Scans Head => • 3-dimensional grid of 12 bit grey values • Depending on resolution: few MB to multiple GB (Example: 2.25 MB) • Image strongly depends on the used scanner and the scan parameters • DICOM header contains some meta information for each slice • DICOM headers are mostly empty or even misleading y Height axis z x <= Feet 23000CWZ8S.0001145710.4.1 1, 16, 31, 46, 61, 76, 91, 106, 121
Picture Archievingand Communication Systems • Scans are stored in Picture Archiving and Communication Systems(PACS) • Scan retrieval by patient name, time, DICOM information • Querying parts of scans is not supported very well=> load and transmit complete scan • No access to sub volumes specified by an example
Problems • slices outside the ROI increase the transfer volume • bottleneck is the LAN: • large transfer times (up to minutes) • bandwidth in LAN is a limiting factor • Only transferring the ROI requires tracking it • slice numbers in the target scan are not the same as in the query scan (scan regions vary) • positions might vary between scans and patients • (organ positions vary)
Region of interest Query CT scan vq User-defined 3D ROI Examplescanvq + chosen ROI Locate ROI Client ID of CT scan vi Target scanvi on remote Server Image database(PACS) Server Trans-fer ROI target scan vi in PACS Client Matching ROI in targetscanvi result
Outline • ROI Queries on CT-Scans • ROI Retrieval Based On a General Height Scale: • Simple Solution based on Similarity Search • Solution based on Generalized Height Scale • kNN Regression for mapping slices • Iterative interpolation • Experimental Validation • Summary
Localization via Similarity Search • Short Commings: • Requires pre-processing or heavy load on server:For each slice in target scan: • Feature Transformation • Comparison to ROI • Feature similarityisinfluencedby global scansimilarity: • Scan parameters • Patient characteristics • => Directsimilarityoftenfails Examplescanvq + chosen ROI Locate ROI Target scanvi on remote Server Trans-fer ROI Matching ROI in targetscanvi
ROI Query Based on a Generalized Height Scale ID of CT scan vi Examplescanvq + chosen ROI Locate ROI CT scan vq User-defined 3D ROI Client H Instance-based Regression Target scanvi on remote Server H gen. height scale H Server Iterative Interpolation Trans-fer ROI CT scan vi in PACS Client Matching ROI in targetscanvi result Height axis (Hscan)
Instance-based Regression: scan→ H Emrich et al: CT Slice Localization via Instance-Based Regression, SPIE‘10 Examplescanvq + chosen ROI Locate ROI H k-NN query: Training Database: 2D imagefeaturesofheight-annotated CT slicesofmultiplescans H Consensus heighth H H Target scanvi on remote Server • Speed-up Measures: • Dimension reduction: RCA • Spatial Indexing: X-Tree • Better: Large training set • Provides multiple examples annotated within consensus height space H • More stable results Bar-Hillel et al: Learning distancefunctions usingequivalencerelations, ICML‘03 Berchtold et al: The X-Tree: An index structure for highdimensional data, VLDB‘96
Iterative InterpolationH→ scan • Combine Regression Mapping with Interpolation vi height space H 3 3 1 1 2 2 CT scan vi (in PACS) 0 H 1 2 3 Estimate location of vi in H via regression Interpolate target positions and . for hlb and hub Verify target positions via regression Accept Result Refinement Interpolation
Outline • ROI Queries on CT-Scans • ROI Retrieval Based On a General Height Scale: • Simple Solution based on Similarity Search • Solution based on Generalized Height Scale • kNN Regression for mapping slices • Iterative interpolation • Experimental Validation • Summary
Quality of Height Regression (scan→ H) Quality andRuntimew.r.t. Training Database size Main Memory Runtimes on original, 175-dimensional Image Features On-disc runtime for dataset of 2103 CT scans (= 0.9 Mio slices) after RCA dimension reduction + X-Tree Indexing: dim 10 => 20 ms With feature generation and dimension reduction: Time / Query = 40 ms Error = 1.98 cm
Validation of ROI Query Pipeline • Testing Height Range Queries on 5 manually-annotated Landmarks in 33 CT Scans: • Annotation Error (LB + UB): 2.6 cm • ROI Query Error (LB + UB): 2.6 – 2.4 cm • ROI Query Runtimes: 1.3 – 10 seconds Pays off if 8 slices are saved sacralpromontory lowerxiphoidprocess lowerplateofthe 12ththoracicvertebra cranialsternum lowerboundofcoccyx Volume oforigin: 23000BVEFR.0000740833.11.1
Runtime Advantages 20 runtime 20 % retrievedslices 15 15 % Runtime [sec] 10 Retrievedfractionofcompletevolumes 10 % 50 5 % 0 0 % Left kidney Urinary bladder Hip to lower L5 Arch of aorta 16.8 cm 9.6 cm 4.7 cm 0.9 cm • Retrieval Times forTypicalQueries: Test on 20 CT scans of 12,000 slices Complete Retrieval time: 70 s per scan => 70 to 99 % reduction of the retrieved volumes
Outline • ROI Queries on CT-Scans • ROI Retrieval Based On a General Height Scale: • Simple Solution based on Similarity Search • Solution based on Generalized Height Scale • kNN Regression for mapping slices • Iterative interpolation • Experimental Validation • Summary
Conclusionand Outlook • Introduced ROI Query Framework: • Great speed-upof CT subvolumeretrievalqueries • Low costsandlowerroroflocalization • Example-basedqueriesare extensible toqueriesusinganatomicalatlases • Future Work: • Extension ofheightqueriestoarbitrary 3D queries • Test on alternative, non-medicalusecases
Backup: Quality of Height Regression (scan→ H) Increased Quality andRuntimewith Database size Main Memory Runtimes RCA dimension reduction + X-Tree Runtimes
Backup: Runtime Advantages • Simulated real-wolrdqueriesofvaryingheights: For 20 CT scans of 12,000 slices: total retrieval time = 1,400 seconds => 70 to 99 % reduction of the retrieved volumes