2023-10-24 16:33:35
The classification, by radiomic algorithms, of nodules not determined in chest CT scan poses a problem within the radiological community. A review of current image analysis approaches for this classification is the subject of a recent article published in the Journal Radiology.
The implementation of low-dose chest computed tomography (CT) (LDCT) for lung screening presents a critical opportunity to advance lung cancer care through early detection. Furthermore, millions of lung nodules are detected incidentally each year, increasing the possibilities for early diagnosis of lung cancer.
Heterogeneous chest CT acquisition practices which do not help to formalize algorithm learning methods
The methodological challenges currently encountered in translating decision aids into clinical practice, as well as technical obstacles related to heterogeneous imaging parameters, optimal feature selection, model choice and the need for sets of well-annotated image data for training and validation purposes, are the subject of specific work, with a view to the ultimate integration of these potentially powerful decision aids into routine clinical practice.
In 2011, the National Lung Screening Trial (NLST) demonstrated a 20% benefit in mortality following three annual LDCT screenings on chest radiography. This result showed that LDCT allows detection of lung cancer at earlier and more treatable stages in people who smoke and are at high risk of lung cancer. Subsequent results from the NELSON (Nederlands – Leuvens Longkanker Screenings Onderzoek) trial and an analysis of 10-year mortality in the Italian Multicenter Lung Screening Trial also confirmed lung cancer mortality benefits of 27% to 39%. % in individuals screened by LDCT.
The tricky case of indeterminate pulmonary nodules on CT
Despite the advantages of LDCT screening in terms of mortality in high-risk asymptomatic populations, the high prevalence of indeterminate non-calcified pulmonary nodules worries specialists, because it leads to high false positive rates. In the NLST, any noncalcified nodule with a maximum transverse diameter of 4 mm or more was considered a positive screening result. Over the three screening cycles, screening positivity in the NLST LDCT arm was 24.2%, but was only 16.8% at final screening because nodules observed to be stable over time might be classified as negative.
The use of CT scanning has increased significantly over time and has resulted in a rapid increase in the detection of indeterminate nodules. Although variously defined, these are generally defined as focal opacities 6 to 30 mm in diameter without clearly benign patterns of calcification or intralesional fat. Determining which ones are malignant among the vast majority that are benign is a critical unmet need that will only increase as the use of CT in routine diagnosis and screening becomes more common.
An article provides a review of current image analysis approaches for the classification of these nodules
Morphological characteristics of nodules are known to influence the likelihood that the nodules are lung cancer, and these characteristics have been incorporated as variables in diagnostic prediction models for lung nodules. Beyond the semantic characterization of indeterminate nodules, advanced image analyzes using radiomics features, machine learning and deep learning algorithms are increasingly used to improve their classification as well as prognosis and optimal management newly diagnosed lung cancers. An article published in the Revue Radiology provides an overview of current image analysis approaches to refine the classification of indeterminate nodules and early lung cancer prognosis in newly diagnosed lung cancers.
Heterogeneity in imaging acquisition and reconstruction significantly affects radiomics and deep learning features. In the absence of convergence of the radiological community on standardized protocols in clinical practice, the challenge is to standardize images before feature extraction or to reduce variations in extracted features. Given the breadth of extractable features, feature selection and dimension reduction techniques are essential for machine learning models to be confidently applied to new datasets.
The radiology community can only adopt these machine learning features and models by demonstrating their adaptation to variations in acquisition protocols, their contributions to knowledge in specific clinical contexts, and their reproducibility in different patient populations. Additionally, it is essential to provide high-quality, publicly available annotated clinical and image datasets that can be used for training and validation purposes. Once all of these conditions are met, as these classification algorithms become understandable and plausible to humans, computer algorithms can pave the way for more personalized approaches to diagnosis and treatment.
Bruno Benque with RSNA
1698166001
#Review #approaches #radiomic #classification #indeterminate #nodules #effect
