AI System Combines CT Imaging with Clinical and Genetic Data for Early Lung Cancer Detection

By MedImaging International staff writers
Posted on 20 Feb 2024

Image: A new study suggests CT imaging with automated AI system can predict EGFR genotype (Photo courtesy of 123RF)

Lung carcinoma prognosis has evolved significantly with the discovery of molecular targets and their corresponding treatments. Specifically, mutations in the Epidermal Growth Factor Receptor (EGFR) gene, found in lung carcinoma, serve as key targets for specialized therapies. However, in countries with limited resources like India, advanced testing methods such as next-generation sequencing remain inaccessible for widespread use. Challenges also include obtaining sufficient tissue from lung core biopsies and dealing with the inherent intratumoral heterogeneity that complicates the identification of suitable tumor tissues. Now, researchers have demonstrated that an AI-based system can automatically detect and analyze lung nodule features from CT images, predicting the likelihood of EGFR mutations. This innovation aids oncologists and patients in resource-limited settings by providing near-optimal care and guiding appropriate treatment decisions.

Previous studies leveraging AI with CT imaging have shown promise in categorizing and analyzing lung nodules without incurring additional costs. However, most of these methods have focused solely on nodule detection in CT images. Moreover, while AI has been used to extract comprehensive lung information for predicting EGFR genotype and evaluating responses to targeted lung cancer therapy, such efforts have predominantly been centered on White and Chinese populations. With a primary focus on the Indian population, researchers led by the Rajiv Gandhi Cancer Institute and Research Centre (New Delhi, India) set out to develop an AI-based strategy that could not only detect but also characterize lung nodules, indicating the EGFR mutational status in lung carcinoma patients. This would help triage patients requiring extensive molecular profiling of the EGFR-driver gene.

The team created a fully automated AI-based Predictive System (AIPS) using machine learning (ML) and deep learning (DL) algorithms. This system can detect lung nodule features from CT images and assess the probability of an EGFR mutation, thus eliminating the need for time-consuming image annotation by radiologists and complex feature engineering. In addition to incorporating EGFR gene sequencing and CT imaging data from 2277 lung carcinoma patients across three cohorts in India and a White population cohort from TCIA, the researchers used the LIDC-IDRI cohort to train the AIPS-Nodule (AIPS-N) model. This model automatically detects and characterizes lung nodules. The AIPS-N model's combination with clinical factors in the AIPS-Mutation (AIPS-M) model was evaluated for its effectiveness in predicting the EGFR genotype, achieving area under the curve (AUC) values ranging from 0.587 to 0.910. The AIPS-N successfully detected nodules with an average AP50 of 70.19% and predicted scores for five lung nodule properties. This research suggests that CT imaging combined with an automated lung-nodule analysis AI system can non-invasively and cost-effectively predict EGFR genotype, identifying patients with EGFR mutations.