Generalization optimizing machine learning to improve CT scan radiomics and assess immune checkpoint inhibitors' response in non-small cell lung cancer: a multicenter cohort study

Background

Recent developments in artificial intelligence suggest that radiomics may represent a promising non-invasive biomarker to predict response to immune checkpoint inhibitors (ICIs). Nevertheless, validation of radiomics algorithms in independent cohorts remains a challenge due to variations in image acquisition and reconstruction. Using radiomics, we investigated the importance of scan normalization as part of a broader machine learning framework to enable model external generalizability to predict ICI response in non-small cell lung cancer (NSCLC) patients across different centers.

Methods

Radiomics features were extracted and compared from 642 advanced NSCLC patients on pre-ICI scans using established open-source PyRadiomics and a proprietary DeepRadiomics deep learning technology. The population was separated into two groups: a discovery cohort of 512 NSCLC patients from three academic centers and a validation cohort that included 130 NSCLC patients from a fourth center. We harmonized images to account for variations in reconstruction kernel, slice thicknesses, and device manufacturers. Multivariable models, evaluated using cross-validation, were used to estimate the predictive value of clinical variables, PD-L1 expression, and PyRadiomics or DeepRadiomics for progression-free survival at 6 months (PFS-6).

Results

The best prognostic factor for PFS-6, excluding radiomics features, was obtained with the combination of Clinical + PD-L1 expression (AUC = 0.66 in the discovery and 0.62 in the validation cohort). Without image harmonization, combining Clinical + PyRadiomics or DeepRadiomics delivered an AUC = 0.69 and 0.69, respectively, in the discovery cohort, but dropped to 0.57 and 0.52, in the validation cohort. This lack of generalizability was consistent with observations in principal component analysis clustered by CT scan parameters. Subsequently, image harmonization eliminated these clusters. The combination of Clinical + DeepRadiomics reached an AUC = 0.67 and 0.63 in the discovery and validation cohort, respectively. Conversely, the combination of Clinical + PyRadiomics failed generalizability validations, with AUC = 0.66 and 0.59.

Conclusion

We demonstrated that a risk prediction model combining Clinical + DeepRadiomics was generalizable following CT scan harmonization and machine learning generalization methods. These results had similar performances to routine oncology practice using Clinical + PD-L1. This study supports the strong potential of radiomics as a future non-invasive strategy to predict ICI response in advanced NSCLC.

Comparison of Measurement Techniques and Response Criteria for MR Imaging Follow-up in Adult Primary Central Nervous System Lymphoma

BACKGROUND AND PURPOSE

Current guidelines proposed for the measurement of primary central nervous system lymphoma in 2005 have indicated that unidimensional and bidimensional measurements may be used, using the same threshold for response categorization, because no clinical study has evaluated the agreement among the measurement techniques. Hence, our study assessed the agreement among different measurements.

MATERIALS AND METHODS

In this retrospective study, primary central nervous system lymphoma lesions were measured with different techniques (longest 1D, axial 1D, 2D, 3D, and the Response Evaluation Criteria in Solid Tumor) on consecutive MR images. Intra- and interobserver correlations were calculated with intraclass correlation coefficients. Correlations between raw measurements and variations in size compared with baseline were evaluated with the Spearman rank correlation, and agreement among response categories was evaluated.

RESULTS

A total of 304 examinations obtained in 40 patients was assessed. The intraobserver intraclass correlation coefficient for 3D, 2D, and longest 1D were ≥0.993. The interobserver intraclass correlation coefficient was ≥0.967. The correlations in raw measurements and size variation in comparison with 3D were respectively; 0.99 and 0.98 for 2D; 0.94 and 0.92 for longest 1D; 0.94 and 0.83 for axial 1D; and 0.90 and 0.79 for Response Evaluation Criteria in Solid Tumor. With 20%-30% and 25%-50% thresholds for unidimensional techniques, response categorizations were 95% and 95% for 2D, 92.5% and 90% for the longest 1D, 87.5% and 82.5% for axial 1D, and 90% and 85% for the Response Evaluation Criteria in Solid Tumor.

CONCLUSIONS

Both longest 1D and 2D demonstrated excellent correlations with 3D measurements. The longest 1D could be used for the follow-up of primary central nervous system lymphoma. If unidimensional measurements were used, 20% and 30% cutoffs should be used for defining response categorization instead of the current guidelines.

Management of Primary Central Nervous System Lymphoma Using Intra-Arterial Chemotherapy With Osmotic Blood-Brain Barrier Disruption: Retrospective Analysis of the Sherbrooke Cohort

Background

Primary central nervous system lymphomas (PCNSL) are rare and aggressive CNS tumors. Current management involves high-dose methotrexate (HD-MTX) typically administered intravenously (IV), despite the existence of the blood-brain barrier (BBB), which significantly decreases its bioavailability. Cerebral intra-arterial chemotherapy (CIAC) coupled with osmotic BBB disruption (OBBBD) can theoretically circumvent this issue.

Methods

We performed a retrospective analysis of patients with newly diagnosed PCNSL treated with HD-MTX-based CIAC+OBBBD at our center between November 1999 and May 2018. OBBBD was achieved using a 25% mannitol intra-arterial infusion. Patients were followed clinically and radiologically every month until death or remission. Demographics, clinical and outcome data were collected from the medical record. All imaging studies were reviewed for evidence of complication and outcome assessment. Kaplan-Meier analyses were used to compute remission, progression-free survival (PFS) as well as overall survival times. Subgroup analyses were performed using the log rank test.

Results

Forty-four patients were included in the cohort. Median follow-up was 38 months. Complete response was achieved in 34 patients (79%) at a median of 7.3 months. Actuarial median survival and PFS were 45 months and 24 months, respectively. Age, ECOG and lesion location did not impact outcome. Complications included thrombocytopenia (39%), neutropenia (20%), anemia (5%), seizures (11%), stroke (2%), and others (20%).

Conclusion

CIAC using HD-MTX-based protocols with OBBBD is a safe and well-tolerated procedure for the management of PCNSL. Our data suggests better PFS and survival outcomes compared to IV protocols with less hematologic toxicity and good tolerability, especially in the elderly.

Tracking the Development of Functional Connectomes for Face Processing

Face processing capacities become more specialized and advanced during development, but neural underpinnings of these processes are not fully understood. The present study applied graph theory-based network analysis to task-negative (resting blocks) and task-positive (viewing faces) functional magnetic resonance imaging data in children (5-17 years) and adults (18-42 years) to test the hypothesis that the development of a specialized network for face processing is driven by task-positive processing (face viewing) more than by task-negative processing (visual fixation) and by both progressive and regressive changes in network properties. Predictive modeling was used to predict age from node-based network properties derived from task-positive and task-negative states in a whole-brain network (WBN) and a canonical face network (FN). The best-fitting model indicated that FN maturation was marked by both progressive and regressive changes in information diffusion (eigenvector centrality) in the task-positive state, with regressive changes outweighing progressive changes. Hence, FN maturation was characterized by reductions in information diffusion potentially reflecting the development of more specialized modules. In contrast, WBN maturation was marked by a balance of progressive and regressive changes in hub-connectivity (betweenness centrality) in the task-negative state. These findings suggest that the development of specialized networks like the FN depends on dynamic developmental changes associated with domain-specific information (e.g., face processing), but maturation of the brain as a whole can be predicted from task-free states.