Limits of Dose Reduction in CT: How low is too low? Acknowledgement. Disclosures 8/2/2012

Transcription

1 8/2/22 Limits of Dose Reduction in CT: How low is too low? Ehsan Samei Duke University Acknowledgement Disclosures Research grant: NIH Research grant: DHS Research grant: RSNA, QIBA Research grant: General Electric Research grant: Siemens Medical Research grant: Carestream Health Share-holder: Zumatek Inc Consultant: NIOSH, GE, Carestream, GLG Council

2 8/2/22 Why do we need to reduce radiation dose? 38 ma 7 ma 8% of dose AJR January 28: 9 3 mas 4 mas* 2 mas* Karmazyn, B et al. AJR January 29 2

3 8/2/22 Qualitative preference performance Image quality changes monotonically with dose 34 Lower ma dose produces lower ¼ of ma IQ How much IQ is good enough? How can we objectively measure IQ? Outlook Imaging provides a clinical benefit: Image quality Universally-appreciated Illusive to define and measure Imaging involves a level of cost Radiation dose: Universally-depreciated Unclear what to reduce and by how much Outlook We cannot optimize what we cannot define or measure We cannot optimize dose in isolation from image quality 3

4 8/2/22 How can we determine the appropriate dose level?. Reasonable measures of radiation burden estimated probability of harm 2. Reasonable measures of image quality needed information for effective clinical management 3. Justified balance btw the two by targeted adjustment of system settings 4. Standardization and consistent implementation of the process What is image quality? Diagnostic accuracy Expensive Impractical given the extent and variability of technologies Image esthetics Subjective and qualitative Not necessarily reflective of diagnostic performance 4

5 8/2/22 What is image quality? Physical metrics Convenient and practical Limited in scope, but Effective if related to diagnostic performance under qualified conditions. CNR 2. Model observer metrics 3. Quantitative metrics CNR Related to detectability for constant resolution and noise texture (Rose, 948) st order approximation of image quality Task-generic CNR = Q L Q B σ B Why CNR is not enough Parameters that affect detectability. Contrast 2. Lesion size 3. Lesion shape 4. Lesion edge profile 5. Noise magnitude 6. Noise texture 7. Resolution 8. Viewing distance 9. Display size Parameters that CNR considers. Contrast 2. Noise magnitude 5

6 8/2/22 Why CNR is not enough IR FBP Solomon, AAPM 22 Noise texture vs kernel GE Siemens Solomon, AAPM 22 Texture similarity Sharpness 6

7 8/2/22 Peak frequency Sharpness Methods to measure detectability Parameters that affect detectability. Contrast 2. Lesion size 3. Lesion shape 4. Lesion edge profile 5. Noise magnitude 6. Noise texture 7. Resolution 8. Viewing distance 9. Display size Task function (W) Noise power spectrum (NPS) Modulation transfer function (MTF) Eye filter (E) Model observers Resolution and contrast transfer Attributes of image feature of interest Image noise magnitude and texture 7

8 /2/22 Model observers Fisher-Hotelling observer (FH) ' ( d FH ) 2 = MTF 2 2 (u,v)w Task (u,v) dudv NPS(u,v) Non-prewhitening observer (NPW) ' [ MTF ( d NPW ) 2 (u,v)w 2 2 Task (u,v) dudv] 2 = MTF 2 2 (u,v)w Task (u,v) NPS(u,v)dudv NPW observer with eye filter (NPWE) ' [ MTF ( d NPWE ) 2 (u,v)w 2 2 Task (u,v) E 2 (u,v)dudv ] 2 = MTF 2 2 (u,v)w Task (u,v) Eye filter NPS(u,v)E 4 (u,v) + MTF 2 2 (u,v)w Task (u,v)n i dudv Spatial frequency (/mm) MTF - axial coronal NPS - axial -.5 coronal ^ Spatial frequency (/mm) MTF 2 - axial coronal NPS 2 - axial coronal Spatial frequency (/mm) ^ Model observers Task functions A Z = f (d ') Task function (AU) Task functions Small feature no iodine Large feature no iodine.. 8 kvp (a) 8 kvp (b).8 kvp.8 kvp 2 kvp 2 kvp 4 kvp 4 kvp kvp 2 kvp.4.4 mm mm Spatial frequency (/mm) Small feature with iodine Task function (AU) kvp kvp 2 kvp 4 kvp 2 kvp mm (c) Spatial frequency (/mm) Large feature with iodine 8 kvp kvp 2 kvp 4 kvp 2 kvp mm (d) Spatial frequency (/mm) Spatial frequency (/mm) 8

9 8/2/22 Validation of model observers NPWE best matched observer performance for FBP, across different doses and tasks Richard, Li, Samei, SPIE 2 Christianson, AAPM 22 Parameter Duke-UMD-NIST study GE Discovery CT 75 HD Siemens Flash Philips ict kvp Rotation time SFOV DFOV Recon Algorithm FBP Standard / ASIR 5 B3f / I3f Safire 3 B / idose5 Recon Mode Helical Helical Helical Collimation Pitch Slice Thickness.25 Dose levels: 2, 2, 7.2, 4.3,.6,.9 mgy For anonymity will be referred to as vendor A, B, and C Observer study design 9

10 8/2/22 Observer study design 3 scanner models 3 dose levels (out of 7) 2 reconstruction algorithms slices x 5 repeated exams Total of 9 images 2 expert observers from two institutions 2AFC with,8 images AUC converted to d CNR vs observer performance d vs observer performance

11 8/2/22 Performance vs image noise, eg IR gives lower performance (AUC) for a given noise level.95 2 kvp Large feature (mm).95 2 kvp Small feature ( mm) AUC SAFIRE -IRIS -FBP AUC Noise (HU) Noise (HU) Performance vs image noise, eg 2 IR gives lower performance (AUC) for a given noise level Large feature (mm) Small feature ( mm) AUC.9.8 -MBIR -ASIR -FBP AUC Noise (HU) Noise (HU) Model observer qualifications Task-based Requires task definition Requires generalization for optimizing task-generic systems Non-linear systems require prescribed evaluation conditions eg, using contrast and noise relevant to the targeted task

12 8/2/22 MTF Resolution and noise, eg Comparable resolution -FBP -IRIS -SAFIRE NPS x -6 Lower noise but different texture -FBP -IRIS -SAFIRE Spatial frequency Spatial frequency Resolution and noise, eg 2 Higher resolution.5 x -5 Lower noise but different texture MTF MBIR -ASIR -FBP NPS Spatial frequency Spatial frequency Mercury Phantom Duke University 2

13 8/2/22 Mercury Phantom Three tapered, four cylindrical regions of polyethelene (Diameters: 6, 23, 3, 37 cm) Cylindrical inserts air, polystyrene, acrylic, teflon different concentration of iodinated targets Wilson, AAPM 22 Winslow, AAPM 22 imquest (image quality evaluation software) HU, Contrast, Noise, CNR, MTF, NPS, and d per patient size, ma modulation profile Wilson, AAPM 22 Winslow, AAPM 22 IQ vs dose 3

14 8/2/22 CNR vs dose 2 ACR Acrylic insert 8 6 CNR Eff dose (msv).2 msv 3.7 msv d vs dose (large feature).95 IR ACR Acrylic insert AUC (NPWE) % 2% FBP Eff dose (msv).2 msv 3.7 msv d' vs dose (small feature).95 IR ACR.5 lp/mm bar pattern AUC (NPWE) % 29% FBP Eff dose (msv).2 msv 3.7 msv 4

15 8/2/22 How much can IR reduce dose? Default clinical protocol at 2 mgy 23% dose reduction How much can IR reduce dose? Default clinical protocol at 2 mgy 6% dose reduction How much can IR reduce dose? Default clinical protocol at 2 mgy 33% dose reduction 5

16 8/2/22 Small s mfeature a ll le s io n n(no o c ocontrast) n t s t Large L a rg feature e le s io n n(no o c ocontrast) n t s t AUC A Z AUC A Z AAUC Z.. f f f f EffE dose d o s e ((msv) m S v ) Eff E dose d o s e ( m(msv) S v ) Small Sfeature m a ll le s io n (with w / c o ncontrast) t s t Large Lfeature a rg e le s io n (with w / c o n contrast) s t. 9 AUC A Z FBP 8 kvp FBP kvp FBP 2 kvp FBP 4 kvp IRIS 8 kvp IRIS kvp IRIS 2 kvp IRIS 4 kvp Eff Edose f f d o s e (msv)c( m S v ) 6 cm EffE dose f f d o s e ((msv)c m S v ) Small feature (no contrast) s m a ll le s io n n o c o n t ra s t To do list: Large L a rg feature e le s io n n(no o c ocontrast) n t s t AUC A Z AUC A Z AAUC Z.. f f f f EffE dose d o s e ((msv) m S v ) Eff E dose d o s e ( m(msv) S v ) Small Sfeature m a ll le s io n (with w / c o ncontrast) t s t Large Lfeature a rg e le s io n (with w / c o n contrast) s t. 9 AUC A Z FBP 8 kvp FBP kvp FBP 2 kvp FBP 4 kvp IRIS 8 kvp IRIS kvp IRIS 2 kvp IRIS 4 kvp Eff Edose f f d o s e (msv)c( m S v ) 24 cm EffE dose f f d o s e ((msv)c m S v ). 9 s m a ll le s io n n o c o n t ra s t To do list: Small feature (no contrast) Large L afeature rg e le s io n n o (no c o n tcontrast) s t. 9 AUC A Z AAUC Z AAUC Z.. f f f f EffE dose d o s e ( m(msv) S v ) E d o s e ( m S v ) Eff dose (msv) Small feature S m a ll le s io n (with w / c o n t racontrast) s t AUC A Z Large feature L a rg e le s io n (with w / c o n tracontrast) s t FBP 8 kvp FBP kvp FBP 2 kvp FBP 4 kvp IRIS 8 kvp IRIS kvp IRIS 2 kvp IRIS 4 kvp Eff Edose f f d o s e (msv)c ( m S v ) 32 cm E f f d o s e ( m S v ) Eff dose (msv) 6

17 8/2/22 Task function (W task ): Virtual model, task characteristics Designer nodule * 2 r C ( r ) = C ( ) n peak R Simulated iodinated lesion Noise free background Burgess-Samei lesion model Task function (W task ): Virtual model, task characteristics Size Iodine concentration Y Lesion diameter = {5, 7.5,, 2.5, 5 mm} Lesion contrast = {, 5, 2, 25, 3 HU} Task function (W task ) d for FBP and IRs FBP d IRs vs. FBP Lesion contrast d' FBP vs. d' IR IRIS vs. FBP SAFIRE3 vs. FBP SAFIRE5 vs. FBP 2 Lesion size 5 5 Lesion size (mm) SAFIRE5 > SAFIRE3 = IRIS > FBP 7

18 8/2/22 AUC vs. dose FBP IRIS SAFIRE3 SAFIRE5 Dose reduction w/ IR vs task AUC threshold =.9; Dose reduction potential (relative value with respect to FBP) for a medium size patient 7% 6% 5% 4% 3% 2% % IRIS SAFIRE3 SAFIRE5 % Quantitative IQ F Chen, SPIE 22 8

19 8/2/22 Li, AAPM 22 Li, AAPM 22 mgy 3 mgy 6 mgy mgy 2 msv.6 per 6 msv. per Li, AAPM 22 9

20 8/2/22 Conclusions Quality is paramount Dose reduction Is possible BUT requires science Requires surrogates of image quality Task- and indication-specific Protocol-specific Patient-specific Requires surrogates of dose Individualized Scalar Cross-modality Conclusions Quality-dose gradient may be used as a basis for dose setting and reduction Dose operating point is a more important target to aim than percent dose reduction Iso-gradient operating points for targeted indications Conclusions Iterative recons and post-processing Significant dose reduction Different noise texture than FBP Resolution dependent on contrast, noise Highly implementation-dependent Effective use requires 2 nd -order image quality metrology 2