Unsupervised machine learning identifies predictive progression markers of IPF.

OBJECTIVES: To identify and evaluate predictive lung imaging markers and their pathways of change during progression of idiopathic pulmonary fibrosis (IPF) from sequential data of an IPF cohort. To test if these imaging markers predict outcome. METHODS: We studied radiological disease progression in 76 patients with IPF, including overall 190 computed tomography (CT) examinations of the chest. An algorithm identified candidates for imaging patterns marking progression by computationally clustering visual CT features. A classification algorithm selected clusters associated with radiological disease progression by testing their value for recognizing the temporal sequence of examinations. This resulted in radiological disease progression signatures, and pathways of lung tissue change accompanying progression observed across the cohort. Finally, we tested if the dynamics of marker patterns predict outcome, and performed an external validation study on a cohort from a different center. RESULTS: Progression marker patterns were identified and exhibited high stability in a repeatability experiment with 20 random sub-cohorts of the overall cohort. The 4 top-ranked progression markers were consistently selected as most informative for progression across all random sub-cohorts. After spatial image registration, local tracking of lung pattern transitions revealed a network of tissue transition pathways from healthy to a sequence of disease tissues. The progression markers were predictive for outcome, and the model achieved comparable results on a replication cohort. CONCLUSIONS: Unsupervised learning can identify radiological disease progression markers that predict outcome. Local tracking of pattern transitions reveals pathways of radiological disease progression from healthy lung tissue through a sequence of diseased tissue types. KEY POINTS: • Unsupervised learning can identify radiological disease progression markers that predict outcome in patients with idiopathic pulmonary fibrosis. • Local tracking of pattern transitions reveals pathways of radiological disease progression from healthy lung tissue through a sequence of diseased tissue types. • The progression markers achieved comparable results on a replication cohort.

Variability of computed tomography radiomics features of fibrosing interstitial lung disease: A test-retest study.

OBJECTIVES: To investigate the intra- and inter-scanner repeatability and reproducibility of CT radiomics features (RF) of fibrosing interstitial lung disease (fILD). METHODS: For this prospective, IRB-approved test-retest study, CT data of sixty fILD patients were acquired. Group A (n = 30) underwent one repeated CT scan on a single scanner. Group B (n = 30) was scanned using two different CT scanners. All CT data were reconstructed using different reconstruction kernels (soft, intermediate, sharp) and slice thicknesses (one and three millimeters), resulting in twelve datasets per patient. Following ROI placement in fibrotic lung tissue, 86 RF were extracted. Intra- and inter-scanner RF repeatability and reproducibility were assessed by calculating intraclass correlation coefficients (ICCs) for corresponding kernels and slice thicknesses, and between lung-specific and non-lung-specific reconstruction parameters. Furthermore, test-retest lung volumes were compared. RESULTS: Test-retest demonstrated a majority of RF is highly repeatable for all reconstruction parameter combinations. Intra-scanner reproducibility was negatively affected by reconstruction kernel changes, and further reduced by slice thickness alterations. Inter-scanner reproducibility was highly variable, reconstruction parameter-specific, and greatest if either soft kernels and three-millimeter slice thickness, or lung-specific reconstruction parameters were used for both scans. Test-retest lung volumes showed no significant difference. CONCLUSION: CT RF of fILD are highly repeatable for constant reconstruction parameters in a single scanner. Intra- and inter-scanner reproducibility are severely impacted by alterations in slice thickness more than reconstruction kernel, and are reconstruction parameter-specific. These findings may facilitate CT data and RF selection and assessment in future fILD radiomics studies collecting data across scanners.

Radiomics score predicts acute respiratory distress syndrome based on the initial CT scan after trauma.

OBJECTIVES: Acute respiratory distress syndrome (ARDS) constitutes a major factor determining the clinical outcome in polytraumatized patients. Early prediction of ARDS is crucial for timely supportive therapy to reduce morbidity and mortality. The objective of this study was to develop and test a machine learning-based method for the early prediction of ARDS derived from the first computed tomography scan of polytraumatized patients after admission to the hospital. MATERIALS AND METHODS: One hundred twenty-three patients (86 male and 37 female, age 41.2 ± 16.4) with an injury severity score (ISS) of 16 or higher (31.9 ± 10.9) were prospectively included and received a CT scan within 1 h after the accident. The lungs, including air pockets and pleural effusions, were automatically segmented using a deep learning-based algorithm. Subsequently, we extracted radiomics features from within the lung and trained an ensemble of gradient boosted trees (GBT) to predict future ARDS. RESULTS: Cross-validated ARDS prediction resulted in an area under the curve (AUC) of 0.79 for the radiomics score compared to 0.66 for ISS, and 0.68 for the abbreviated injury score of the thorax (AIS-thorax). Prediction using the radiomics score yielded an f1-score of 0.70 compared to 0.53 for ISS and 0.57 for AIS-thorax. The radiomics score achieved a sensitivity and specificity of 0.80 and 0.76. CONCLUSIONS: This study proposes a radiomics-based algorithm for the prediction of ARDS in polytraumatized patients at the time of admission to hospital with an accuracy that competes and surpasses conventional scores despite the heterogeneous, and therefore more realistic, scanning protocols. KEY POINTS: • Early prediction of acute respiratory distress syndrome in polytraumatized patients is possible, even when using heterogenous data. • Radiomics-based prediction resulted in an area under the curve of 0.79 compared to 0.66 for the injury severity score, and 0.68 for the abbreviated injury score of the thorax. • Highlighting the most relevant lung regions for prediction facilitates the understanding of machine learning-based prediction.

[Machine learning in radiology : Terminology from individual timepoint to trajectory]. / Maschinelles Lernen in der Radiologie : Begriffsbestimmung vom Einzelzeitpunkt bis zur Trajektorie.

METHODICAL ISSUE: Machine learning (ML) algorithms have an increasingly relevant role in radiology tackling tasks such as the automatic detection and segmentation of diagnosis-relevant markers, the quantification of progression and response, and their prediction in individual patients. STANDARD RADIOLOGICAL METHODS: ML algorithms are relevant for all image acquisition techniques from computed tomography (CT) and magnetic resonance imaging (MRI) to ultrasound. However, different modalities result in different challenges with respect to standardization and variability. METHODICAL INNOVATIONS: ML algorithms are increasingly able to analyze longitudinal data for the training of prediction models. This is relevant since it enables the use of comprehensive information for predicting individual progression and response, and the associated support of treatment decisions by ML models. PERFORMANCE: The quality of detection and segmentation algorithms of lesions has reached an acceptable level in several areas. The accuracy of prediction models is still increasing, but is dependent on the availability of representative training data. ACHIEVEMENTS: The development of ML algorithms in radiology is progressing although many solutions are still at a validation stage. It is accompanied by a parallel and increasingly interlinked development of basic methods and techniques which will gradually be put into practice in radiology. PRACTICAL CONSIDERATIONS: Two factors will impact the relevance of ML in radiological practice: the thorough validation of algorithms and solutions, and the creation of representative diverse data for the training and validation in a realistic context.

Radiomic features define risk and are linked to DNA methylation attributes in primary CNS lymphoma.

Background: The prognostic roles of clinical and laboratory markers have been exploited to model risk in patients with primary CNS lymphoma, but these approaches do not fully explain the observed variation in outcome. To date, neuroimaging or molecular information is not used. The aim of this study was to determine the utility of radiomic features to capture clinically relevant phenotypes, and to link those to molecular profiles for enhanced risk stratification. Methods: In this retrospective study, we investigated 133 patients across 9 sites in Austria (2005-2018) and an external validation site in South Korea (44 patients, 2013-2016). We used T1-weighted contrast-enhanced MRI and an L1-norm regularized Cox proportional hazard model to derive a radiomic risk score. We integrated radiomic features with DNA methylation profiles using machine learning-based prediction, and validated the most relevant biological associations in tissues and cell lines. Results: The radiomic risk score, consisting of 20 mostly textural features, was a strong and independent predictor of survival (multivariate hazard ratio = 6.56 [3.64-11.81]) that remained valid in the external validation cohort. Radiomic features captured gene regulatory differences such as in BCL6 binding activity, which was put forth as testable treatment target for a subset of patients. Conclusions: The radiomic risk score was a robust and complementary predictor of survival and reflected characteristics in underlying DNA methylation patterns. Leveraging imaging phenotypes to assess risk and inform epigenetic treatment targets provides a concept on which to advance prognostic modeling and precision therapy for this aggressive cancer.

Dynamic memory to alleviate catastrophic forgetting in continual learning with medical imaging.

Medical imaging is a central part of clinical diagnosis and treatment guidance. Machine learning has increasingly gained relevance because it captures features of disease and treatment response that are relevant for therapeutic decision-making. In clinical practice, the continuous progress of image acquisition technology or diagnostic procedures, the diversity of scanners, and evolving imaging protocols hamper the utility of machine learning, as prediction accuracy on new data deteriorates, or models become outdated due to these domain shifts. We propose a continual learning approach to deal with such domain shifts occurring at unknown time points. We adapt models to emerging variations in a continuous data stream while counteracting catastrophic forgetting. A dynamic memory enables rehearsal on a subset of diverse training data to mitigate forgetting while enabling models to expand to new domains. The technique balances memory by detecting pseudo-domains, representing different style clusters within the data stream. Evaluation of two different tasks, cardiac segmentation in magnetic resonance imaging and lung nodule detection in computed tomography, demonstrate a consistent advantage of the method.

Prospects and Challenges of Radiomics by Using Nononcologic Routine Chest CT.

Chest CT scans are one of the most common medical imaging procedures. The automatic extraction and quantification of imaging features may help in diagnosis, prognosis of, or treatment decision in cardiovascular, pulmonary, and metabolic diseases. However, an adequate sample size as a statistical necessity for radiomics studies is often difficult to achieve in prospective trials. By exploiting imaging data from clinical routine, a much larger amount of data could be used than in clinical trials. Still, there is only little literature on the implementation of radiomics in clinical routine chest CT scans. Reasons are heterogeneous CT scanning protocols and the resulting technical variability (eg, different slice thicknesses, reconstruction kernels or timings after contrast material administration) in routine CT imaging data. This review summarizes the recent state of the art of studies aiming to develop quantifiable imaging biomarkers at chest CT, such as for osteoporosis, chronic obstructive pulmonary disease, interstitial lung disease, and coronary artery disease. This review explains solutions to overcome heterogeneity in routine data such as the use of imaging repositories, the standardization of radiomic features, algorithmic approaches to improve feature stability, test-retest studies, and the evolution of deep learning for modeling radiomics features. Supplemental material is available for this article. © RSNA, 2020 See also the commentary by Kay in this issue.

Automatic lung segmentation in routine imaging is primarily a data diversity problem, not a methodology problem.

BACKGROUND: Automated segmentation of anatomical structures is a crucial step in image analysis. For lung segmentation in computed tomography, a variety of approaches exists, involving sophisticated pipelines trained and validated on different datasets. However, the clinical applicability of these approaches across diseases remains limited. METHODS: We compared four generic deep learning approaches trained on various datasets and two readily available lung segmentation algorithms. We performed evaluation on routine imaging data with more than six different disease patterns and three published data sets. RESULTS: Using different deep learning approaches, mean Dice similarity coefficients (DSCs) on test datasets varied not over 0.02. When trained on a diverse routine dataset (n = 36), a standard approach (U-net) yields a higher DSC (0.97 ± 0.05) compared to training on public datasets such as the Lung Tissue Research Consortium (0.94 ± 0.13, p = 0.024) or Anatomy 3 (0.92 ± 0.15, p = 0.001). Trained on routine data (n = 231) covering multiple diseases, U-net compared to reference methods yields a DSC of 0.98 ± 0.03 versus 0.94 ± 0.12 (p = 0.024). CONCLUSIONS: The accuracy and reliability of lung segmentation algorithms on demanding cases primarily relies on the diversity of the training data, highlighting the importance of data diversity compared to model choice. Efforts in developing new datasets and providing trained models to the public are critical. By releasing the trained model under General Public License 3.0, we aim to foster research on lung diseases by providing a readily available tool for segmentation of pathological lungs.

Volumetry based biomarker speed of growth: Quantifying the change of total tumor volume in whole-body magnetic resonance imaging over time improves risk stratification of smoldering multiple myeloma patients.

The purpose of this study was to improve risk stratification of smoldering multiple myeloma patients, introducing new 3D-volumetry based imaging biomarkers derived from whole-body MRI. Two-hundred twenty whole-body MRIs from 63 patients with smoldering multiple myeloma were retrospectively analyzed and all focal lesions >5mm were manually segmented for volume quantification. The imaging biomarkers total tumor volume, speed of growth (development of the total tumor volume over time), number of focal lesions, development of the number of focal lesions over time and the recent imaging biomarker '>1 focal lesion' of the International Myeloma Working Group were compared, taking 2-year progression rate, sensitivity and false positive rate into account. Speed of growth, using a cutoff of 114mm3/month, was able to isolate a high-risk group with a 2-year progression rate of 82.5%. Additionally, it showed by far the highest sensitivity in this study and in comparison to other biomarkers in the literature, detecting 63.2% of patients who progress within 2 years. Furthermore, its false positive rate (8.7%) was much lower compared to the recent imaging biomarker '>1 focal lesion' of the International Myeloma Working Group. Therefore, speed of growth is the preferable imaging biomarker for risk stratification of smoldering multiple myeloma patients.

Effects of individualized electrical impedance tomography and image reconstruction settings upon the assessment of regional ventilation distribution: Comparison to 4-dimensional computed tomography in a porcine model.

Electrical impedance tomography (EIT) is a promising imaging technique for bedside monitoring of lung function. It is easily applicable, cheap and requires no ionizing radiation, but clinical interpretation of EIT-images is still not standardized. One of the reasons for this is the ill-posed nature of EIT, allowing a range of possible images to be produced-rather than a single explicit solution. Thus, to further advance the EIT technology for clinical application, thorough examinations of EIT-image reconstruction settings-i.e., mathematical parameters and addition of a priori (e.g., anatomical) information-is essential. In the present work, regional ventilation distribution profiles derived from different EIT finite-element reconstruction models and settings (for GREIT and Gauss Newton) were compared to regional aeration profiles assessed by the gold-standard of 4-dimensional computed tomography (4DCT) by calculating the root mean squared error (RMSE). Specifically, non-individualized reconstruction models (based on circular and averaged thoracic contours) and individualized reconstruction models (based on true thoracic contours) were compared. Our results suggest that GREIT with noise figure of 0.15 and non-uniform background works best for the assessment of regional ventilation distribution by EIT, as verified versus 4DCT. Furthermore, the RMSE of anteroposterior ventilation profiles decreased from 2.53±0.62% to 1.67±0.49% while correlation increased from 0.77 to 0.89 after embedding anatomical information into the reconstruction models. In conclusion, the present work reveals that anatomically enhanced EIT-image reconstruction is superior to non-individualized reconstruction models, but further investigations in humans, so as to standardize reconstruction settings, is warranted.

Heterogeneity and matching of ventilation and perfusion within anatomical lung units in rats.

Prior studies exploring the spatial distributions of ventilation and perfusion have partitioned the lung into discrete regions not constrained by anatomical boundaries and may blur regional differences in perfusion and ventilation. To characterize the anatomical heterogeneity of regional ventilation and perfusion, we administered fluorescent microspheres to mark regional ventilation and perfusion in five Sprague-Dawley rats and then using highly automated computer algorithms, partitioned the lungs into regions defined by anatomical structures identified in the images. The anatomical regions ranged in size from the near-acinar to the lobar level. Ventilation and perfusion were well correlated at the smallest anatomical level. Perfusion and ventilation heterogeneity were relatively less in rats compared to data previously published in larger animals. The more uniform distributions may be due to a smaller gravitational gradient and/or the fewer number of generations in the distribution trees before reaching the level of gas exchange, making regional matching of ventilation and perfusion less extensive in small animals.