Download presentation
Presentation is loading. Please wait.
Published byJanessa Caulder Modified over 10 years ago
1
Biomediq Biomedical Image Quantification www.biomediq.com Biomediq Personalized Breast cancer Screening Mads Nielsen
2
Biomediq Image Group
3
Biomediq Image Group
4
Biomediq Image Group
5
Biomediq Image Group
6
Biomediq Image Group
7
Biomediq Image Group
8
Biomediq Image Group
9
Biomediq Motivation and summary In the western world women are mammography screened for breast cancer differentiated on age only using more than 100.000.000 mammograms per year. Risk profiling may be used to personalize screening frequency and/or technology to reduce cost and increase detection rate. Breast density may provide an essential risk profiling tool. We present the Breast Cancer Risk Meter based on density and mammographic breast tissue texture doubling the risk segregation compared to breast density alone. Patented technology is using the screening mammogram to automatically asses the 4 year risk of breast cancer.
10
Biomediq Participants Biomediq A/S, Mads Nielsen Management Image analysis Prototype development DIKU, Chrstian Igel Learning on massive data Online learning Capitol Region Screening Program, Ilse Vejborg Data collection Clinical testing Department of Public Helath, Elsebeth Lynge Risk modeling Health economics
11
Biomediq BIRADS Examples of the 4 categories 1234
12
Biomediq Method: BC Risk Meter Assess local pixel image structures [1] Match these to database of structures with known outcome (4y BC diagnosis) [2] Decide category for each pixel (e.g. healthy/4y BC diagnosis) Integrate local decisions to global score of the image Combine this with age and density into a risk estimate 1 2
13
Biomediq Local Image Structure Local Image Features
14
Biomediq Method 3: Examples of density scores LowMediumHigh
15
Biomediq Clinical studies overview S1: Purpose: to demonstrate risk segregation ability Materials: 245 cases/250 controls form the Dutch screening program. Digitized film 2-4 year prior to diagnosis. Result: risk segregation capability adds to percentage density S2: Purpose: to independently verify sensitivity and demonstrate robustness Materials: 226 cases/442 controls form the Minnesota Cohort. Digitized film 3-6 year prior to diagnosis Result: adds robustly a factor of two to risk segregation odds ratio S3: Purpose: to demonstrate robustness to modality and estrogen receptor status 145 cases/423 controls from the Pennsylvania Cohort. Direct Digital Mammography on contralateral breast including estrogen receptor status. Result: verifies robustness to modality and shows relation to ER status
16
Biomediq S1: Breast cancer risk study * Non-parametric MethodAUCp* 2BI-RADS0.58< 10 -2 3Percentage0.60< 10 -4 4BC MTR0.63< 10 -8 5Aggregate 3+40.66< 10 -12 245 cancer cases (123 interval cancers and 122 screen detected) and 250 age- matched controls from the Dutch breast cancer screening program Area under the ROC curve seperating cancers and controls
17
Biomediq S2: Breast cancer validation study S2: Mammograms 3-6 years prior to diagnosis Controls(N=442)Cases(N=226) P-value Age54.8± 10.555.8± 10.60.28 BMI27.9± 6.627.9± 5.50.96 PD18.4± 14.722.0± 15.40.003 S1: Mammograms 2-4 years prior to diagnosis Controls(N=250)Cases(N=245) P-value Age61.3± 6.461.7± 8.80.19 PD13.2± 10.216.9± 11.1<0.001 Adjusted for BMI, Age, Menopause, HRT PD0.61 BC MTR0.60 PD + BC MTR0.66 Result of recognizing texture from S1 in S2
18
Biomediq S3: Direct Digital Mammography Examination at time of diagnosis, but from the unaffected mammogram. 111 ER+ cases, 34 ER- cases, 423 controls. Differentiation of receptor status is based on recognition of texture in mammograms of known status. Cancer vs control: Separation of ER+ and ER- AUCp-value PD0.56<0.05 BC MTR0.64<0.001 PD + BC MTR0.65<0.001 AUCp-value PD0.51NS BC MTR0.64<0.05 PD + BC MTR0.70<0.001
19
Biomediq Simple simulation of BC MTR personalization of screening PD is used for referral to high risk programs. PD is published to be expedient for screening frequency personalization. These simulation are based on S1 and the 4 year prognosis instead of the relevant 2 year prognosis Whenever 100,000 women are screened using a recall rate of app. 2% -462 cancers will be detected -189 cancers will not be detected but show before next round By excluding 20 % of the women with lowest score in next round -16 cancers extra will show as interval cancers -This is less than half the rate as in normal screening Hence, the women not screened in next round will have the half false negative rate compared to whole population now. By referring 10% to high risk analysis -42 % of the interval cancers will be in this group
20
Biomediq Breast cancer technical details New coordinate system with anatomical orientation
21
Biomediq Breast cancer technical details Features used are: - 3 jet - Horizontal heterogenity at scales 1 mm, 2mm, 4mm, 8mm - Posisition kNN-classification, k=100 20 X 1000 random points per image, and SFS of features = 20 committees Fusion of 20.000 x 100 NN by average
22
Biomediq Project content Retrospective study of all 2011 cancers at all time points 8 x 4 x 2 x 360 images Prospective study of all women screened in 2012-2015 2 x 4 x 160,000 images of 80 Mb each Machine learning on massive data Building a prototype
23
Biomediq Prototype
24
Biomediq Project conclusion Will this technology make it possible to increase the early detection rate and save ressources at the same time?
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.