There is a need for open-source software that integrates AI/ML algorithms into clinical workflows for pre-clinical evaluation [6]. Jansen et al., [9] developed a vendor-agnostic platform for integrating AI applications into digital pathology infrastructures, and Sohn et al., [13] introduced a vendor-agnostic platform for integrating AI into radiology infrastructures using breast density on 2D mammography as a use case. XNAT-OHIF integrates with DICOM [4] but to our knowledge and based on personal correspondence, did not provide quantitative volumetric information. Segmentation and quantification of pathology— such as advanced malignancy [14], lung nodule size [15], or COVID infiltrate volume [16] has generated considerable interest as potential precision medicine tools since manual segmentations are not feasible at the point of care, and there is considerable information loss and subjectivity associated with diameter-based measurements [17].
In this work, we address an unmet need for tools that integrate automated cross-sectional imaging segmentation results into a DICOM-based quantitative visualization clinical workflow. Since its introduction in 2021 [3], nnU-net has emerged as a widely-employed robust and easy to train method for segmentation tasks in medical imaging in the NIfTI format. Our open-source vendor-agnostic software is intended for clinical-translational researchers who wish to deploy their segmentation models in inference for further testing on new cases encountered in the clinical workflow using the DICOM standard. For those wishing to use cascaded nnU-net, the pipeline can be used out-of-the-box with relevant .pkl files.
On the virtual infrastructure host side, a router/listener anonymizes and handles DICOM series which are sent to a DICOM query/retrieve archive backing an OHIF web viewer, and to an on-premises single GPU-based DL workstation. On the DL host side, DICOM series are converted to NIfTI and processed by the segmentation algorithm. A NIfTI segmentation mask sharing the same UID as the DICOM files is converted to a DICOM SEG object and returned to the router/listener where a DICOM SR element containing segmentation volume (in mL) is created. The DICOM SEG and SR objects are then sent to the DICOM archive for viewing. The segmentation and quantitative information are thereby harmonized to the same format as the original DICOM data. The building blocks were implemented using publicly available open-source libraries, which made our software vendor-agnostic and easily deployable, along FAIR principles. By open-sourcing the proposed software, we encourage radiologists and radiology IT developers to integrate more data transfer functionality and more models into the clinical radiology workflow.
Radiologists should be able to receive verifiable quantitative results well within CT report turnaround times should they wish, for example, to include this information in their reports within the framework of a prospective research study. We tested the software using 21 consecutive patients with traumatic pelvic hematoma. Clinical interpretation of WBCT scans for polytrauma or cancer staging typically exceeds 30 minutes, and results were available within a fraction of this minimum expected turnaround time in all cases.
Using our method, we achieved a mean clock time of 5 minutes and 32 seconds using a workstation with a single NVIDIA GeForce RTX 3090 Ti graphics card. This is approximately 1/5th of a typical report turnaround time for a patient undergoing WBCT for suspected polytrauma. nnU-net inference is responsible for over 89% of the clock time, and the time for all other on-premises DL host-side and virtual router/listener-side steps were found to be negligible, with a mean of only 38.5 seconds (which includes the 30 second time-out). We surmise that investigators will encounter similar or shorter clock times for less complex use cases.
There are limitations to our pilot study. We describe clock times for 21 patients on a single task. However, any algorithm or model can be used. We include a publicly available nnU-net model for spleen segmentation (pretrained nnU-net model Task009_Spleen) in our GitHub link to initially operationalize the deployed pipeline. In the future, end-users may wish to have an "always-on” system that sends the series of interest for every patient directly from a scanner AE Title. Given the lag time associated with post-processing, study completion by the technologist, and transfer from the scanner to PACS, sending a given series from the scanner on creation could result in substantial time savings, however this may not be desirable without an initial rapid detection or classification step to separate positive from negative studies for a given feature of interest. We are currently working on these steps for our problem and plan to release future updates. Sending a study from PACS to the listener/router node selected from a drop-down menu is currently the only manual step. To simplify the process, we are working on an integrated PACS icon. We are also exploring solutions for pop-up notifications and auto-population of quantitative results in radiology reports. Our method currently employs nnU-net and investigators wishing to implement other segmentation algorithms and models will need to make minor modifications to our code.
In conclusion, we have developed and released a simple open-source vendor-agnostic PACS and DICOM compatible software package for automated quantitative visualization with nnU-net. The method, intended to promote “shadow evaluation” using new cases in the clinical setting, approximates FDA-designated IPQ or CADx quantitative volumetry-based CAD tools, and is meant to help advance the application of precision medicine principles for cross-sectional imaging.
Tables