At SemanticMD, we’ve been developing a platform to facilitate the acquisition and curation of medical imaging and clinical data to develop artificial intelligence (AI) algorithms. As a team of machine learning scientists ourselves, we realized that curation of training data is often the rate limiting factor for developing new AI algorithms and image analysis techniques.
In order to solve the problem for medical imaging researchers, we developed a cloud-based platform and suite of software tools to facilitate the development of AI algorithms, including tools to redact images, annotate reports using NLP, crowd-source annotations from medical imaging specialists, and APIs for developers to quickly test their favorite deep learning networks. Our solution is specialized for medical imaging and enables seamless data curation within clinical workflows for machine learning data set generation.
Acquire, Store, & View
SemanticMD Data Market
Find, connect and license medical imaging data with expert annotations. The first step in the process is getting access to sufficient data. We have data partners globally with access to radiology (PET, CT, US, MR, X-ray), ophthalmology (fundus, OCT) and dermatology (dermoscopy and whole body) data. In addition, we have expert annotators on staff that can provide annotations such as patient diagnoses, nodule marking, and lesion segmentation. A common use case for machine learning teams is to license cases for algorithm validation and use our annotator team to clean/correct algorithm predictions.
Spend less time loading and transferring images and more time getting things done. Organize your medical images, create reports, and view any medical image in one place. SemanticMD Storage is the perfect solution for researchers looking for an AI-enabled teaching filesystem or medical imaging startups looking for an affordable storage backend. We support DICOM, HL7, and REST API connectivity. SemanticMD Storage and associated viewers can handle DICOM, JPG/PNG, videos and even gigapixel microscopy images.
Annotate & Redact
Automate your data collection. Use natural language processing (NLP) to annotate radiological and clinical reports for search and analysis. PACS solutions aren’t typically designed for medical imaging researchers. If you were trying to create an algorithm to detect hemorrhagic stroke automatically, the first step would be gathering relevant cases. While a PACS might let you find all the Head CT cases, you would still have to go through all the reports to classify cases as being positive or negative for hemorrhagic stroke. With NLP you could quickly categorize all the data in your PACS by ICD-10/SNOMED diagnoses. Recently, the NIH took this approach to creating a database of 100K+ CXR scans. We SemanticMD NLP you can curate a large machine learning dataset quickly and with minimal time and resources.
Collaborate and annotate faster. SemanticMD Annotate enables your team to organize medical image annotation projects in a fun, flexible way and output the results to JSON for easy analysis by machine learning teams. Sometimes categorizing the study isn’t enough and you need more granular annotations across multiple sources to establish ground truth. We offer the first crowd-sourced platform designed for medical imaging data. Annotate your data at the study level, series level, and image/instance level all with an easy to use user management system.
Find and remove protected health information (PHI) in text and images. In addition to removing PHI from DICOM headers, we also handle text burned into medical images/videos and redacting dermatology images. We provide easy APIs that you can integrate into your own systems for anonymization and redaction.
Compute & Deploy
Train and manage deep learning models in a beautifully simple and visual app. SemanticMD Compute enables training and testing of state-of-the-art deep learning models specific to medical imaging. Once you have your annotated data, the next step in a typical AI project is to start testing out different algorithms and gauging initial performance. Our dashboard for managing deep learning projects gives you access to cloud GPUs and a set of APIs to train and test your favorite deep learning networks (e.g. GoogleNet, ENet, SegNet, YOLO, etc.). Get in touch with us for the complete list of networks and pre-processing algorithms supported by SemanticMD Compute. Being able to quickly test different algorithms on your annotated datasets is crucial to deciding next steps on how to improve performance. Next steps may include annotating more data, refining pre-processing steps, or modifying network architecture.
Once you have a trained and validated deep learning algorithm, the final step is deployment. We provide easy APIs, packaged VMs (e.g. Docker), and custom deployment services (e.g. on device integration). In addition, our model compression and translation services ensure that your team gets an algorithm you can use in real-time in the cloud or on device.