Smartphone Imaging Repository: A Novel, Cost-Effective Method For Creating a CT Image Bank

and but Typical We describe and test an image capture method using smartphone camera video images of brain computed tomography (CT) scans of traumatic intracranial hemorrhage. The deidentied videos are emailed from the emergency department for central adjudication. We selected eight scans, mild moderate and severe subdural and multicompartmental hematomas and mild and moderate intraparenchymal hematomas. Eighty videos were made by 10 people on 7 different smartphones. We measured the time in seconds it took to capture and email the les. The primary outcomes were hematoma volume measured by ABC/2, Marshal Scale, midline shift measurement, image quality by contrast-to-noise ratio (CNR) and time to capture. A radiologist and an imaging scientist applied ABC/2 method, calculated the Marshall scale and midline shift on the video images and on the PACS in a randomized order. We calculate the intraclass correlation coecient (ICC). We measured image quality by calculating contrast-to-noise ratio (CNR). We report summary statistics on time to capture in the smartphone group without a comparator. Analyzed in multiple ways -- image quality, contrast to noise ratio, calculation of volume, SDH thickness, midline shift and Marshal Score – smartphone scans were not inferior to PACS read scans. They can be collected and emailed in about a minute.


Introduction
Imaging repositories are valuable tools in understanding diseases but assembling them can be costly, even when attaching them to an ongoing clinical trial or drawing them from existing clinical medical records. Many clinical research sites still use compact discs to store images and mail them to trial sponsors. Electronic transfer is becoming more common 1 but is still billed at 1 to 2 hours of staff time per scan. Staff often must physically go to the radiology suite to either write the CD or use other means of transmission. Smartphones have revolutionized the incorporation of imaging for consultation of specialists, making obsolete a century of lexicon in describing wounds or fractures. Clinicians routinely send deidenti ed images of X-rays or lacerations from their own phones or those provided by the hospital. Smartphone cameras have improved so much that entire feature-length lms are shot with them 2 . Applying this pocket technology to imaging has been happening in the clinical arena for years, but it has not yet been tested as a method for assembling an image repository.

Study Overview
Ethics This study has regulatory and ethical approvals for waiver of consent. All data were deidenti ed and delinked from patient records. The shift-T function on this PACS system removes all text, which removes all identi ers.

Design
Cohort study comparing PACS read CT images to smartphone captured and processed images.

Setting
CT scans from the trauma registry at a Level 1 trauma center in Austin, Texas.

Image Cohort Selection
We selected 8 standard of care baseline computed tomography (CT) scans from traumatic intracranial hemorrhage patients in the trauma registry to represent mild, moderate and severe subdural hematomas and multicompartmental hemorrhage, and mild and moderate isolated intraparenchymal contusions. Subdural severity degree was divided by 0-5 mm width, 5-10 mm width and >10 mm width. Multicompartmental and intraparenchymal hemorrhage was de ned as total hematoma volume of all bleeding sites of 1-10 cc, 10-30 cc and >30 cc. All volumes were calculated by the simpli ed ABC/2 method 3,4 . The axial CT images were acquired from the skull base to the vertex without contrast. Dose lowering techniques included adjusting the mA and/or kV per standard of care.

Variables
The primary outcome variables were hematoma volume measured by ABC/2, Marshal Scale, midline shift measurement, image quality by contrast-to-noise ratio (CNR) and time to capture. A radiologist and an imaging scientist applied ABC/2 method, calculated the Marshall scale and midline shift on the video images and on the PACS in a randomized order. We calculate the intraclass correlation coe cient (ICC).
We measured image quality by calculating contrast-to-noise ratio (CNR). We report summary statistics on time to capture in the smartphone group without a comparator.

Data Source
Trauma registry standard of care CT scans of traumatic intracranial hemorrhage.

Bias
We selected the scans to represent different severity of disease, which is known to affect ability to measure and is thus a better test of a new capture method, but this could be seen as a selection bias. The raters were unblinded to the study purpose which could introduce bias.

Data Collection
Data collectors included a convenience sample of emergency department (ED) staff present on a Saturday morning, including 3 attending emergency physicians, 2 emergency resident physicians, 1 emergency nurse, 1 emergency technician, 2 social workers and a clerk. A central ED radiology reading station was used to access the Picture Archive and Communication System (PACS) via Synapse version 4.4 (Fuji lm Healthcare, Lexington, MA). The images were in DICOM format with text (and thus identi ers) hidden. Each data collector used a smartphone to record the screen while using the arrow button to scroll through axial slices of the non-contrast CT of each selected scan. The smartphones used were 3 iPhone 12s, 2 iPhone 11s, 1 iPhone 10, 1 iPhone 9, a Samsung Galaxy S7, a Samsung Galaxy S9 and an Android OnePlus. Data collection methods are outlined in Figure 1.

Image Data Management
Videos were saved in mp4 or mov format at native resolution and emailed via a secure healthcare email platform for post processing. The resulting les were evaluated for completeness, technical quality, and adherence to the protocol requirements. Video les were converted to tagged image le (tif) format at a rate of 30 frames per second using Python version 3.7.2. Image analyses were performed using ImageJ 5 a Java-based image processing program developed at the National Institutes of Health and the Laboratory for Optical and Computational Instrumentation. All CT images were randomly presented to avoid observer recall bias.

Image Data Collection
Image read data were collected and managed using REDCap (Research Electronic Data Capture) electronic data capture tools hosted at Dell Medical School at The University of Texas Austin 6,7 . The embedded scalebar in each deidenti ed image was used to calibrate the scale and obtain data values that related to spatial resolution including subdural hematoma thickness, midline shift, and hematoma volume (via the ABC/2 method) 3,4 . To quantitatively assess lesion conspicuity, image contrast to noise was calculated via [1] where SI indicates the mean signal intensity within either the lesion or surrounding tissue and σ is the standard deviation of the image noise.
The Marshall score for diffuse head injury was recorded for each data set 8 . The ratings included: 1, Normal for age; 2, High/mixed density mass less than 25cc, midline shift less than 5mm, basilar cisterns preserved; 3, Basilar cisterns effaced & High/mixed density mass less than 25cc, midline shift less than 5mm, basilar cisterns preserved; 4, Midline shift greater than 5mm; 5, Evacuated mass lesion: High/mixed density mass >25cc which was surgically evacuated; 6, Non-evacuated mass lesion: High/mixed density mass >25cc not surgically treated.

Image Raters
The image raters were one imaging scientist and one radiologist working independently and reviewing the images in a randomized order, but we could not blind them as the PACS images were read on PACS and the smartphone images were read on ImageJ.

Primary Outcomes
The primary outcomes were hematoma volume measured by ABC/2, Marshal Scale, midline shift measurement, image quality by contrast-to-noise ratio (CNR) and time to capture.

Statistics
All statistical analyses were completed in RStudio version 1.4.1106 (PBC, Boston, MA) using R version 4.1.0 (https://www.R-project.org/). Interrater reliability was evaluated using the intraclass correlation coe cient (ICC). Speci cally, the ICC1 was calculated to re ect the amount of total variance of the measurements for each lesion type in relation to the total variance using the Pearson correlation with an = 0.05. To further assess the level of agreement between image analysis on the PACS vs. smartphones for quantitative measures (volume, SDH thickness, and midline shift), Bland Altman analyses were performed with acceptable limits de ned a priori to ensure the calculated critical difference was greater than the difference between each smartphone measurement and PACS.
To detect differences across the multiple phones and users in capturing the quantitative measures and ordinal Marshall and Likert scores for each lesion type, the non-parametric Friedman λ 2 test was used with post-hoc Kendall W to determine effect size. Signi cance for all analyses was set at ∝=0.05.

Results
Example results from the PACS, and each smartphone model are shown in Fig. 2 (Table 2). Further analysis of volume measures (Friedman λ 2 of 9.14, P = .1035, Kendall's W effect size = 0.229) and Marshall score (Friedman λ 2 of 5.56, P = .3519, Kendall's W effect size = 0.139) ensured values derived from different smartphones are not signi cantly different from those measured using the PACS, (Fig. 2) and Bland-Altman plots comparing each phone type to the PACS measures (Fig.   3) support these ndings.

Discussion
Analyzed in multiple ways --image quality, contrast to noise ratio, calculation of hematoma volume, SDH thickness, midline shift and Marshall Score -smartphone scans were not inferior to PACS read scans. A 1,000-patient imaging repository assembled with this method would take about 17 hours of site staff time with images directly emailed to the clinical coordinating center. This is a tiny fraction of the effort using older methods, even with electronic transfers, for which coordinating centers budget 1-2 hours per scan (125-250 8-hour days for 1,000 scans). It does not require emergency or other investigators or their research staff to physically go to the radiology department to collect and transmit images.
We studied deidenti ed, delinked scans for this analysis, but obviously the scans in a repository must be linked to patient data to be useful. We worked with a smartphone application company, RNDS, to solve this problem. RNDS was created by trauma surgeons to be a central HIPAA-compliant cloud repository for patient data along with functionalities for VOIP, video, and SMS, staff-to-staff and staff-to-patient. We adapted this technology for clinical trial purposes. Opening the smartphone camera through the app keeps the images in a HIPAA-compliant environment as the video is streamed into the cloud without saving it to the phone, and the deidenti ed le is assigned the subject's study number, which is linked to that consented patient's clinical data. This real-time approach would allow for quality check software to be installed in the app on the phone, and it creates an opportunity to build tools to check for trial eligibility. It also creates the possibility of simpler methods 9 for a "meta-repository" of multiple clinical trials. We will apply these innovations to a randomized trial of timing of restarting direct oral anticoagulants after traumatic intracranial hemorrhage (Restart TICrH NCT04229758) 10 .
Traumatic intracranial hemorrhage is an ideal disease in which to rst study this method, as intracranial blood on CT is fairly obvious. It remains to be tested whether this approach is viable for more subtle ndings such as early stroke, e.g., loss of gray-white border. Hematoma volume measurements are an important surrogate outcome 11 , and clinical trial site measurements show high variability and low correlation with centrally measured scans 12 . Trials struggle collecting images for central measurement because of mislabeling, scans lost in the mail, failure to deidentify completely, etc. DICOM imaging les commonly contain metadata that is potentially identifying and di cult to remove. Our method collected more than 100 scans in less than 2 hours with multiple operators and phones, and only one was mislabeled, and none contained identifying information. Reading of imaging on smartphones has been previously reported 13 , but there are no reports of capturing images for an imaging repository with this method.

Limitations
We could not blind the readers to whether the images were from smartphone or PACS, as the formats differ. This might introduce bias, though probably more so against the non-traditional scans. Saving the PACS images in a video le potentially would degrade image quality, making it no longer the gold standard comparator. We are exploring converting the video into stills and loading them into the PACS for a blinded comparison. We could not test every smartphone that might be used during the trial. We tested eight different traumatic intracranial hemorrhage scans, which cannot represent the full range of this disease. All Marshall score ratings by the radiologist had perfect agreement on the Marshall score while four reads from the imaging scientist, less familiar with the scoring system, disagreed. There may be scans on the borderline of the Marshall scale that would need the better resolution to make a de nitive classi cation. The same may be true for mixed density hematomas assessed for hematoma volume. It is probably true that most scans do not need this degree of resolution to be classi ed. For those that do, traditional methods could be used to obtain the DICOM images.

Declarations
This study had ethics approval under a natural history protocol for patients with stroke or at risk for stroke by the University of Texas Health Sciences Institutional Review Board.  Figure 1 Work ow for data collection including multiple smartphone operators electronically transferring video les for post-processing and analyses. Variables collected include hematoma volume, subdural hematoma thickness, midline shift, contrast-to-noise ratio, and image quality. Nonparametric evaluation of differences across methods was performed.

Figure 2
Example images of severe multi-compartment lesion including convex subdural and intraparenchymal hematomas derived from the PACS and multiple smartphone video captures.