Mammography limitations. Clinical performance of digital breast tomosynthesis compared to digital mammography: blinded multi-reader study

Transcription

1 Clinical performance of digital breast tomosynthesis compared to digital mammography: blinded multi-reader study G. Gennaro (1), A. Toledano (2), E. Baldan (1), E. Bezzon (1), C. di Maggio (1), M. La Grassa (1), L. Pescarini (1), I. Polico (1), A. Proietti (1), A. Toffoli (1) (1) Oncological Institute of Veneto, I.R.C.C.S., Padova - Italy (2) Biostatistics Consulting, LLC, Toronto - Canada Mammography limitations SCREENING MAMMOGRAPHY = BEST METHOD FOR EARLY DETECTION OF BREAST CANCER FALSE NEGATIVE RATE = 1-3 Retrospective studies = 67 of cancers were visible on the prior mammograms MISSED CANCERS DETECTION ERRORS CHARACTERIZATION ERRORS Inter-observer variability k statistics SLIGHT FAIR MODERATE SUBSTANTIAL ALMOST PERFECT Lazarus et al BI-RADS lexicon for US and mammography: interobserver variability and positive predictive value Radiology 26, 239: radiologists probability of malignancy BI-RADS classification BIRADS 3rd EDITION : 5 classes, increasing with probability of malignancy; BIRADS 4 (suspicious) includes most of probability distribution (5-8 ) a 4b 4c BIRADS "ETENDED" BIRADS BIRADS (ACR Third Edition) Classification accuracy can be improved by dividing BIRADS 4 category into 3 subclasses, corresponding to different risks of malignancy. probability of malignancy "Extended" BIRADS BIRADS 4th EDIITON: 7 classes, increasing with probability of malignancy 1

2 BI-RADS: breast density BI-RADS: masses ACR Description Mostly fatty Fibroglandular Heterogeneously dense Dense Diagnostic accuracy Very high High Limited Limited SHAPE MARGINS DENSITY Obenauer et al, Eur Radiol 25, 15: A LESION CAN BE CONSIDERED A MASS IF IT IS VISIBLE IN BOTH CC AND MLO VIEWS BI-RADS: calcifications The ROC paradigm Obenauer et al, Eur Radiol 25, 15: benign Non-diseased Threshold Diseased concern high probability of malignancy Test result value or subjective judgement of likelihood that case is diseased 2

3 The ROC paradigm Non-diseased Non-diseased Diseased more typically: Diseased TPF, sensitivity less aggressive mindset Test result value or subjective judgement of likelihood that case is diseased FPF, 1-specificity Non-diseased Non-diseased Diseased Threshold TPF, sensitivity moderate mindset Diseased Threshold TPF, sensitivity more aggressive mindset FPF, 1-specificity FPF, 1-specificity 3

4 Non-diseased Entire ROC curve Entire ROC curve Threshold Diseased TPF, sensitivity TPF, sensitivity chance line Reader Skill and/or Level of Technology FPF, 1-specificity FPF, 1-specificity Area under ROC curve (AUC) Area under ROC curve (AUC) 1 1 Overall measure of test performance Comparisons between two tests based on differences between (estimated) AUC True Positive 1 AUC = 1 False Positive 1 True Positive 1 AUC = 5 False Positive 1 True Positive AUC = 9 True Positive AUC = 65 False Positive 1 False Positive 1 4

5 Background Anatomic Noise Background Principles uniform background breast structure Burgess A E et al Lesion detection in digital mammograms Proc SPIEE 21, 4321: Kotre C J The effect of background structure on the detection of low contrast objects in mammography BJR 1998, 71: < LESION DETECTABILITY < CLINICAL PERFORMANCE (sensitivity( sensitivity, specificity) & CTBI: accuracy Gong et al A computer simulation study comparing lesion detection accuracy with digital mammography, breast tomosynthesis, anc cone-beam CT breast imaging Med Phys 26, 33: digital mammography tomosynthesis breast CT Potential of & CTBI WILL DIGITAL BREAST TOMOSYNTHESIS [OR BREAST CT] REPLACE SCREENING MAMMOGRAPHY? [Dr. Dan Kopans, MGH] DETECTION:SIGNIFICANT BENEFIT 5

6 Purpose Method: study population COMPARE CLINICAL PERFORMANCE OF ONE-VIEW DIGITAL BREAST TOMOSYNTHESIS () VERSUS TWO-VIEWS FULL-FIELD FIELD DIGITAL MAMMOGRAPHY () (1 view) (2 views) 2 PATIENTS, US CLINICAL DECISION INCLUSION/ ECLUSION CRITERIA? CONSENT? Work-up (FNAC, VABB, FU, etc.) YES NO STOP TRUTH MLO VS. CC + MLO INCLUSION CRITERIA >= 4 y lesion BIRADS >=3 ( or US) breast size to fit detector FOV (19 x 23 cm2) ECLUSION CRITERIA previous breast mastectomy breast implant high genetic risk Acquisition systems a b c RMLO Case # 69 - projections RMLO 1 Proj. 1/15 Angle = -65 Proj. 8/15 Angle = -45 Proj. 15/15 Angle = -25 *GE investigational device CsI/a /a-si flat panel; 19 x 23 cm 2 ; 1x1 μm 2 pixel size; Mo/Mo, Mo/Rh Rh, Rh/Rh; Manual exposure mode; 15 projections per breast; 4 arc; MLO only. GE Senographe 2D CsI/a /a-si flat panel; 19 x 23 cm 2 ; 1x1 μm 2 pixel size; Mo/Mo, Mo/Rh Rh, Rh/Rh; AOP/STD CC + MLO 15 slice 73/73 slice 46/73 6

7 dose reconstruction: slabs vs. slices Dose MLO < = SFM Dose AGD (mgy) (MLO) screen/film (CC+MLO) breast thickness (mm) Dose CC+MLO N slices 1 SLAB = 1 SLICES slab thickness overlap dose (MLO view) per each breast thickness is compared with 2x Dose Acceptance Limits proposed by the "European Guidelines for quality assurance in breast cancer screening and diagnosis, 4 th edition", derived from screen/film mammography Case # 12 - LMLO Case # 12 - vs (slabs) LMLO slice 8 slice 9 slice 1 slice 11 slice 12 slab [8-12] 7

8 Case # 81 - vs Method: Reading Protocol LMLO LMLO LMLO Method: Reading Protocol : CC+MLO BIRADS (7 steps) TRUTH slice 25/61 slice 44/61 Method: Reading Protocol Method: Reading Protocol slabs slices BIRADS (7 steps) Breast density (BIRADS 1-4) 1 Localization of findings (max 3) Finding Conspicuity (1-5) BIRADS (7-steps) Lesion type Lesion size (mm) slabs slices Most useful view CC MLO SLABS SLICES TRUTH 8

9 Methods Truth establishment Malignant lesions: histology (from surgical or core biopsy); Benign lesions: histology (in case of biopsy), FNAC and/or long/term follow-up (long-term >= 1y history). Negative (no lesion): information from the patient folder or consensus meeting (in case of disagreement) True Positive Fraction First results: clinical performance Reader A AUC =.929 AUC =.728 p =.3 Readers 3 Patients 5 Effective Reader B False Positive Fraction AUC =.852 AUC =.757 p = Reader C AUC =.846 AUC =.797 p =.522 Presented at First results: image quality First conclusions CONSPICUITY SCALE: (subjective) 1. Conspicuity > Conspicuity # lesions not visible low conspicuity high conspicuity Reader A Reader B Reader C NotDetected Low Conspicuity High Conspicuity NotDetected Low Conspicuity High Conspicuity Conspicuity > Conspicuity 3 1 NotDetected Low Conspicuity High Conspicuity 2. AUC > AUC 3. Significant difference for 1 of 3 readers 4. Sample size!!! RESULTS ARE ENCOURAGING AND SUPPORT THE POTENTIAL BENEFIT OF TOMOSYNTHESIS OVER 2D-MAMMOGRAPHY Presented at 9

10 Second step Clinical performance: MRMC ROC Average over 6 Readers - 1 patients Readers 3 6 Patients 5 1 Effective SAME READING PROTOCOL True Positive Fraction AUC =.8434 AUC =.8749 AUC p = False Positive Fraction NOT SIGNIFICANTLY DIFFERENT AT 95 CL Presented at Inter-reader reader variability Second conclusions Readers - 1 patients 1. Clinical performance of (MLO) was slightly superior vs. (CC+MLO), even if not statistically significant; kappa fair agreement poor Overall No Lesion Benign Malignant SIGNIFICANTLY BETTER FOR 2. Inter-reader reader variability was lower with vs. for malignant lesions. RESULTS SUPPORT THE OPPORTUNITY FOR TOMOSYNTHESIS TO REDUCE INTER-READER READER VARIABILITY IN AREAS UNDER ROC CURVES AND IN BIRADS SCORES FOR MALIGNANT LESIONS Presented at 1

11 Third step (final) Readers 6 6 Patients 1 2 Effective SAME READING PROTOCOL Final results presented at Method: Multi-Reader Multi-Case ROC Multiple Readers: : 6 breast radiologists (5 Population: : 2 patients (5-3 y experience) Independent readings of left & right breasts = 371 effective Multiple Reading Sessions: including 5 & 5 images Bias Control: NO & images of the same breast in the same session Results: MRMC ROC analysis Malignant lesions vs. all other breasts Results: MRMC ROC analysis All lesions vs. normal breasts Average over 6 Readers Average over 6 Readers TPF FPF AUC A B C D E F ALL Radiologists TPF FPF AUC A B C D E F ALL Radiologists NOT SIGNIFICANTLY DIFFERENT AT 95 CL VARIABILITY IN AUC s ACROSS READERS SLIGHTLY LOWER FOR (SD=.247) VS. FOR (SD=.426) NOT SIGNIFICANTLY DIFFERENT AT 95 CL 11

12 Results: MRMC ROC analysis Results: non-inferiority mean AUC Malignant Malignant + Benign DIFFERENCE IN AUCs FOR MALIGNANT VS. ALL LESIONS IS HIGHER FOR THAN FOR Mean AUC difference (-) CL +95 CL -95 CL -95 CL non-inferior inferior Malignant Malignant + Benign Breasts with Lesions Conclusions 1. Overall clinical performance with (MLO) was not significantly different vs. (CC+MLO); 2. Higher difference in AUCs for malignant vs. all lesions suggests that could allow radiologists to better discriminate between malignant and benign findings. TOMOSYNTHESIS (1-VIEW) HAS SHOWN TO BE NON-INFERIOR TO DIGITAL MAMMOGRAPHY (2-VIEWS) Perspectives 1. SCREENING: WILL REPLACE MAMMOGRAPHY? Non-inferiority is insufficient (dose/cost-effectiveness) effectiveness) Workflow needs to be proven Some kind of benefit should be proven (ex. drastic reduction in recall rate relevant in Europe?) 2. DIAGNOSTIC: MIGHT BE USEFUL AS AN ADJUNCT TO MAMMOGRAPHY? Retrospective analysis on subset of data to investigate specific indications for (dense breasts, architectural distortions, etc.) Ensure that the same additional information cannot be easily obtained by other non-irradiating / less expensive modalities (US or 2-D 2 D extra-views). 12

13 Thank you for your attention! 13