2023 North America Conference on Lung Cancer (NACLC) - Posters and Abstracts-PP01.26 (Poster) Proprietary vs Open-Source Radiomic Platform for Lung Cancer Diagnosis

Pdf Summary

A study aimed to compare the diagnostic performances of two radiomic feature extractors, one proprietary (HealthMyne) and one open-source (PyRadiomics), in the diagnosis of lung cancer in indeterminate pulmonary nodules (IPNs). IPNs are lung nodules with unclear malignancy potential found incidentally or on screening imaging. Currently, there are various platforms available for risk stratification, but no head-to-head comparisons have been done. <br /><br />The study used a previously validated radiomic signature from HealthMyne as the proprietary comparator. Three radiomic signatures were developed using the open-source PyRadiomics feature extractor with different methods of feature selection. The radiomic models were trained on an internal cohort and externally validated on three cohorts. The diagnostic accuracy was measured by the Area Under the Receiver Operating Characteristic Curve (AUC). <br /><br />Additionally, clinical radiomic (ClinRad) models were developed by combining the Mayo model clinical risk score with the best performing open-source signature and the proprietary signature using a LASSO regression methodology. The diagnostic accuracies of the ClinRad models were evaluated by their AUCs and clinical improvements by their bias-corrected clinical net reclassification indices (cNRIs).<br /><br />The results showed that both the open and proprietary radiomic models improved the diagnostic accuracy of IPNs compared to the Mayo clinical risk score. The open-source platform performed nearly identically to the commercial platform in terms of diagnostic accuracy and improvement in reclassification of nodule malignancy risks. The study supports the use of more accessible, open-source radiomic platforms for analysis in IPNs and encourages the creation and use of open-sourced platforms for research in radiomics.<br /><br />In terms of specific results, the proprietary radiomic signature achieved an AUC of 0.76, while the open-source platform achieved a similar AUC of 0.75. When Mayo scores were added to the open-source LASSO models, forming the ClinRad models, the AUC was identical to the proprietary ClinRad model at 0.80. Both ClinRad models showed clinical improvement compared to Mayo, as measured by the bias-adjusted clinical net reclassification index (cNRI).<br /><br />Overall, this study highlights the comparable performance of open-source radiomic platforms in the diagnosis of lung cancer and supports their use in furthering the field of radiomics.

Asset Subtitle

David Xiao

Keywords

radiomic feature extractors

HealthMyne

PyRadiomics

lung cancer

indeterminate pulmonary nodules

diagnostic accuracy

AUC

clinical radiomic models

Mayo model clinical risk score

open-source radiomic platforms