Artificial intelligence-ready skin cancer alchemy: transforming routine teledermatology data into metadata-embedded DICOM files

Most skin artificial intelligence (AI) classifiers are trained only on images with diagnostic labels. However, the addition of clinical information can improve predictive accuracy. Recent interest has been stimulated in incorporating clinical data into image files, using the well-established international Digital Imaging and Communication in Medicine (DICOM) standards (Caffery L, Weber J, Kurtansky N et al. DICOM in dermoscopic research: experience report and a way forward. J Digit Imaging 2021; 34: 967–73). We have developed an automated process of creating metadata-embedded DICOM files, directly from a live teledermatology system, described below. Through our Community and Locality Imaging Centre (CLIC) model, patients referred from primary care are triaged to CLIC for high-quality image capture. There, trained health professionals use a mobile application to capture standardized DICOM information for each lesion. Each lesion dataset contains images (macroscopic, dermoscopic) and clinical metadata (patient and lesion information). Datasets are transferred to an image management system, for teledermatology and verification of ground-truth diagnoses by a consultant dermatologist. On completion of diagnoses, datasets are flagged for conversion into DICOM format, where metadata are embedded in the image files. Flagged datasets are cleaned and clinical metadata are mapped to DICOM attributes. Datasets are converted into metadata-embedded DICOM files, and reviewed for conformance to the DICOM standard using the open-source fo-dicom library (v5). These files are further tested for conformance to DICOM standard using the dciodvfy validator tool. Compliant DICOM files are then transferred to a trusted research environment for research. To test whether these DICOM files are usable for AI research, they are examined using the DICOM viewing software 3D Slicer (https://www.slicer.org/), ensuring images are usable and metadata are correctly translated. Image pixel data and clinical metadata are extracted using pydicom, into a format suitable for AI algorithm development. In our pilot work, 658 lesion datasets have been converted into metadata-embedded DICOM files. Conversion on existing hardware [virtual Intel central processing units with 2.60 GHz (two processors) and 8 GB of memory] took < 1 s per image. Metadata-embedded DICOM files were approximately 0.2 kB bigger than the original JPEG files. For 3-MB images, this represented a negligible 0.003% increase in storage requirement. Testing has shown that these files can be successfully handled by algorithms within an AI research environment. In summary, we have demonstrated the feasibility of automating the conversion of routine teledermatology data into AI-ready image files encoded with clinical metadata. Future work is planned to evaluate the utility of this output on the performance of AI classifiers.