eeDAP Registration Accuracy Qi Gong, Benjamin P Berman, Marios A Gavrielides, Brandon D. Gallas Division of Imaging, Diagnostics and Software Reliability (DIDSR) Office of Science and Engineering Laboratories (OSEL) Center for Devices and Radiological Health (CDRH) U.S. Food and Drug Administration (FDA) Hello everyone, My name is Qi Gong. I want to update the talk I gave in Pittsburg: eeDAP Registration Accuracy

Outline Recap of preliminary study (first presented at Pathology Informatics) eeDAP Introduction Registration Study Purpose Methods Results Future Plans I will recap our preliminary study, which includes introduction, study purpose methods and results. Then I will show our future plans and hope to get feedback

eeDAP evaluation environment for Digital and Analog Pathology Control the microscope stage so that a user can read the same tissue/cell from WSI and glass slide. Camera view Camera WSI image Introduction: eeDAP stands for evaluation environment for Digital and Analog Pathology. It can control the microscope stage so that a user can read same position of tissue from WSI and glass slide. This is our system. We have a window to show WSI image. The stage controller can control this stage to move slide under microscope. The camera recodes same view as eyepiece and displays it on screen. Stage Stage controller

Registration Accuracy Study Objective: Evaluate the registration accuracy of eeDAP using real tissue. What is the registration error (distance) for each FOV? Does registration accuracy depend on tissue? Registration is very important, Because the end user relies on it to view same area on optical and digital image. For development, the purpose of this study is Evaluating eeDAP registration accuracy using real tissue. We want to find out what is the registration error (distance) for each FOV? And does the accuracy depends on tissue

Study Methods 1. The center of each FOV should be a small target: a corner of cell or tiny structure. In the study, we choose FOV with a small target in the center, like a corner of cell or tiny structure. Fiducial mark : points out center feature.

Study Methods 2. In WSI, use virtual reticle to identify the target. And in WSI image, we use virtual reticle to locate the target.

Study methods 3. On microscope, use reticle in eyepiece Rotatable ruler Measure distance: target to center (human) 1 2 3 After registration, On microscope, reader uses a rotatable ruler eyepiece reticle to measure the distance between target and center of view. 4 5 6 7 8 9 10

Reticle Choice Klarmann Rulings: KR-207 1 2 3 4 5 6 7 8 9 10 Klarmann Rulings: KR-207 250 um in 100 divisions at 40x on specimen plane (rotatable) This is the reticle we use. Under 40x, it has a 250 micro meter ruler. Each small tick is 2.5 micro meter

Local RE-Registration eeDAP microscope mode always starts with global registration During data collection, for each FOV … No local RE-registration: Measure stage motion accuracy, don’t do any local RE-registration. Automatic registration: Focus may not be good. Use center patch of camera image. Fast registration: Focus microscope. Best registration: Use entire camera image. For each FOV there are 4 RE-Registration options, System does the Automatic Registration, when the microscope arrives at the task ROI. It only use then center area of camera image. In this registration the microscope may not be focused well. If the Automatic Registration fails, administrator can focus microscope and do Fast Registration. It also use the center area of camera image. If the Fast Registration fails again, administrator can do the Best Registration, which Use entire camera image. Administrator can use joystick to visually register camera image and WSI

Local RE-Registration eeDAP microscope mode always starts with global registration During data collection, for each FOV … No local RE-registration: Measure stage motion accuracy, don’t do any local RE-registration. Automatic registration: Focus may not be good. Use center patch of camera image. Fast registration: Focus microscope. Best registration: Use entire camera image. Not evaluated Not evaluated For each FOV there are 4 RE-Registration options, System does the Automatic Registration, when the microscope arrives at the task ROI. It only use then center area of camera image. In this registration the microscope may not be focused well. If the Automatic Registration fails, administrator can focus microscope and do Fast Registration. It also use the center area of camera image. If the Fast Registration fails again, administrator can do the Best Registration, which Use entire camera image. Administrator can use joystick to visually register camera image and WSI 2 readers, 3 slides 10 ROIs/slide 1 reader, 3 slides 10 ROIs/slide

Slide choice Canine oral melanoma, phosphohistone H3 (pHH3) stain. Canine oral melanoma, hematoxylin and eosin (H&E) stain. Human lymph node, hematoxylin and eosin (H&E) stain. (challenging case due to uneven tissue: out of focused scanned WSI) 1 2 3 We used these 3 slides to do the study. The first one is Canine melanoma with pHH3 stain? The second one is same tissue with H and E stain? The third is Human lymph node with H and E stain. It is challenging case due to uneven tissue. We want to use them to show different stain and tissue influence. 2mm 2mm 4mm

Human H&E (challenging slide) Results: Reader 1 Fast registration 27 out of 30 FOV “Successes” Successful registration: error less than 2.5 µm (first tick mark of ruler, diameter of cell in the study is about 8 µm). 3 out of 30 FOV “Failures” When it fails, it fails badly: observed error large (between 100 and 150 µm). Best registration Fix 2 out of 3 cases. The failed case error was 75 µm. Fast Registration Best Registration Canine H&E 10 Canine pHH3 9 Human H&E (challenging slide) 8 This is first reader study results In fast registration, 27 out of 30 FOV Succeeded. As the table shows, all canine H and E FOVs succeeded, we got 1 Failure in Canine pHH3, and 2 in the challenging slide. In succeed cases, the errors were less than 2.5 micro meter it is smaller than a cell size. But for Failures, the errors were between 100 to 150 micro meter. The best registration fixed the Canine pHH3 failed FOV and 1 Human H and E failed FOV. But we still got one failed case, the error is 75 micro meter. Outlier ROI shown on slide 10. 100 microns from center feature for both readers? (seems one is more off than the other)

Results: Reader 2 Independent global registration. Independent focus for Fast Registration. 100% matching results with first reader. The second reader did independent global registration and fast registration. He got 100% matching results with the first reader. Outlier ROI shown on slide 10. 100 microns from center feature for both readers? (seems one is more off than the other)

Future Plans Now, we need to expand and complete this study. All registration modes (No, Auto, Fast, Best). Additional tissues? Additional stains? Current plan: 6 slides How many ROI per slide? 10? Across slides versus within one slide. How many readers? 2? Any comments, concerns, suggestions are appreciated. Thank you H&E Canine Lymph Node Breast pHH3 ER HER2 In future, We plan to evaluate accuracy of automatic registration and no RE-registration, that means only use global registration result and don’t do any re-registration. We also plan to extend the study to more stain, tissue, reader and site. We will also try to improve registration methods