Automatic trending and analysis of SPECT quality assurance with artificial intelligence optical character recognition.

Background: To guarantee high-quality patient scans, thorough quality assurance (QA) of SPECT or gamma cameras, including performance, review, and documentation, is essential.

Purpose: We developed a novel Nuclear Medicine Quality Assurance server (NMQA) with an AI deep learning (AIDL) optical character recognition (OCR) system to automate QA data retrieval and review from SPECT and gamma cameras. The system extracts and compares daily and weekly QA data against specifications. Our goal was to improve the efficiency of QA reviews and facilitate trending, storage, and auditing of QA data across our large hospital network.

Methods: The NMQA Server was implemented in a Linux system using open-source Python as the programming language, DICOM tool kit DCMTK for query of QA data, and Pydicom for managing DICOM images and Structured Query Language (SQL) for interacting with a relational MySQL database. The MySQL database stores numerical results for intrinsic and extrinsic floods, MHR, and COR, along with pointers to the image database facilitating trending analysis of numerical values and flood data evaluation. It also streamlines the review through the server's web interface, accessible on iPhones, iPads, and computers. The AIDL OCR is structured into three stages: feature extraction, sequence labeling, and transcription. The OCR comprises two steps: region of interest (ROI) extraction and character recognition. The AIDL OCR was benchmarked for both accuracy and speed against four common OCRs of Tesseract, OCRopus, PhotoOCR, and EasyOCR on a QA dataset, consisting of 60 flood and 6 COR images without post-processing, and evaluated for accuracy on 3459 flood-scans with post-processing.

Results: The new NMQA server can automatically query QA data, avoid the frequent mistake of typographical errors in naming the QA data, extract the numerical values of the QA data, and build a QA database for trending and analysis of the QA data. It takes about 3 min to complete a query of QA data from all 14 scanners and subsequent postprocessing. The web design facilitated review of flood images over days. The time to review the QA data on PACS without the NMQA server was about 60 min and has been reduced to several minutes using the new NMQA server web page on iPhones, iPads, or computers. The AIDL OCR outperformed Tesseract, OCRopus, PhotoOCR, and EasyOCR in speed and accuracy, maintaining CPU-friendly performance with a processing speed of just 0.3 s per image and accuracy of 93.53%. The AIDL OCR achieved an accuracy of 99.9% in recognizing numerical values in the Arial font, with sizes ranging from 10 to 14, specific to the two different kinds of scanners utilized in this study.

Conclusion: The NMQA server automatically queries QA data to avoid the frequent mistake of typographical errors in naming the QA data, eliminates manual checking of the numerical values against the manufacturers' specifications, improves the efficiency of review of the daily flood images and weekly bar resolution phantom images, enables trending and analysis of the QA data for quality assurance and improvement, and documents the QA data and review for auditing.