Machine learning in quantitative PET: A review of attenuation correction and low-count image reconstruction methods

The purpose of this article is to review the status and limitations of anatomic tumor response metrics including the World Health Organization (WHO) criteria, the Response Evaluation Criteria in Solid Tumors (RECIST), and RECIST 1.1. This article also reviews qualitative and quantitative approaches to metabolic tumor response assessment with (18)F-FDG PET and proposes a draft framework for PET Response Criteria in Solid Tumors (PERCIST), version 1.0. PubMed searches, including searches for the terms RECIST, positron, WHO, FDG, cancer (including specific types), treatment response, region of interest, and derivative references, were performed. Abstracts and articles judged most relevant to the goals of this report were reviewed with emphasis on limitations and strengths of the anatomic and PET approaches to treatment response assessment. On the basis of these data and the authors' experience, draft criteria were formulated for PET tumor response to treatment. Approximately 3,000 potentially relevant references were screened. Anatomic imaging alone using standard WHO, RECIST, and RECIST 1.1 criteria is widely applied but still has limitations in response assessments. For example, despite effective treatment, changes in tumor size can be minimal in tumors such as lymphomas, sarcoma, hepatomas, mesothelioma, and gastrointestinal stromal tumor. CT tumor density, contrast enhancement, or MRI characteristics appear more informative than size but are not yet routinely applied. RECIST criteria may show progression of tumor more slowly than WHO criteria. RECIST 1.1 criteria (assessing a maximum of 5 tumor foci, vs. 10 in RECIST) result in a higher complete response rate than the original RECIST criteria, at least in lymph nodes. Variability appears greater in assessing progression than in assessing response. Qualitative and quantitative approaches to (18)F-FDG PET response assessment have been applied and require a consistent PET methodology to allow quantitative assessments. Statistically significant changes in tumor standardized uptake value (SUV) occur in careful test-retest studies of high-SUV tumors, with a change of 20% in SUV of a region 1 cm or larger in diameter; however, medically relevant beneficial changes are often associated with a 30% or greater decline. The more extensive the therapy, the greater the decline in SUV with most effective treatments. Important components of the proposed PERCIST criteria include assessing normal reference tissue values in a 3-cm-diameter region of interest in the liver, using a consistent PET protocol, using a fixed small region of interest about 1 cm(3) in volume (1.2-cm diameter) in the most active region of metabolically active tumors to minimize statistical variability, assessing tumor size, treating SUV lean measurements in the 1 (up to 5 optional) most metabolically active tumor focus as a continuous variable, requiring a 30% decline in SUV for "response," and deferring to RECIST 1.1 in cases that do not have (18)F-FDG avidity or are technically unsuitable. Criteria to define progression of tumor-absent new lesions are uncertain but are proposed. Anatomic imaging alone using standard WHO, RECIST, and RECIST 1.1 criteria have limitations, particularly in assessing the activity of newer cancer therapies that stabilize disease, whereas (18)F-FDG PET appears particularly valuable in such cases. The proposed PERCIST 1.0 criteria should serve as a starting point for use in clinical trials and in structured quantitative clinical reporting. Undoubtedly, subsequent revisions and enhancements will be required as validation studies are undertaken in varying diseases and treatments.

Machine learning is a technique for recognizing patterns that can be applied to medical images. Although it is a powerful tool that can help in rendering medical diagnoses, it can be misapplied. Machine learning typically begins with the machine learning algorithm system computing the image features that are believed to be of importance in making the prediction or diagnosis of interest. The machine learning algorithm system then identifies the best combination of these image features for classifying the image or computing some metric for the given image region. There are several methods that can be used, each with different strengths and weaknesses. There are open-source versions of most of these machine learning methods that make them easy to try and apply to images. Several metrics for measuring the performance of an algorithm exist; however, one must be aware of the possible associated pitfalls that can result in misleading metrics. More recently, deep learning has started to be used; this method has the benefit that it does not require image feature identification and calculation as a first step; rather, features are identified as part of the learning process. Machine learning has been used in medical imaging and will have a greater influence in the future. Those working in medical imaging must be aware of how machine learning works. ©RSNA, 2017.