2023 World Conference on Lung Cancer (Posters)-EP11.03. Quantification of Interstitial Lung Abnormality Using Deep Learning in Cancer Patients Who Underwent Immune Checkpoint Inhibitors

EP11.03. Quantification of Interstitial Lung Abnormality Using Deep Learning in Cancer Patients Who Underwent Immune Checkpoint Inhibitors - PDF(Slides)

Back to course

Pdf Summary

This study aimed to develop an automated segmentation model for quantifying interstitial lung abnormalities (ILA) on chest CT images in cancer patients who underwent immune checkpoint inhibitors (ICIs). ILA refers to subtle interstitial changes in the lung parenchyma. The researchers used a deep learning model called Mask2Former to segment ILA and evaluated its performance using the mean average precision (mAP) and dice similarity coefficient (DSC). Among 291 cancer patients, 27.1% had ILA. The segmentation model achieved an mAP of 27.4% and a DSC of 54.0% for lung cancer patients, and an mAP of 19.9% and a DSC of 49.2% for non-lung cancer patients. The results showed that the model has potential clinical utility in detecting ILA, although further improvements are needed. ILA has been associated with immune checkpoint inhibitor-induced pneumonitis in cancer patients, and its presence is an independent predictor of mortality. The study highlights the importance of accurate and reproducible evaluation of ILA in oncologic patients, which can be facilitated by AI automatic quantification. This study contributes to the development of a deep learning model for ILA segmentation, which can aid in the early detection and monitoring of ILA in cancer patients undergoing ICIs. However, the model's performance is affected by the challenging radiologic features of ILA, resulting in reduced mAP and DSC. The study was conducted retrospectively using a single-center dataset and further validation is required on larger and more diverse cohorts.

Asset Subtitle

Eun Young Kim

Meta Tag

Speaker Eun Young Kim

Topic Metastatic NSCLC: Immunotherapy - Retrospective

Keywords

automated segmentation model

interstitial lung abnormalities

chest CT images

cancer patients

immune checkpoint inhibitors

deep learning model

mean average precision

dice similarity coefficient

clinical utility

early detection