257MO - Integrating AI and ML with lung cancer diagnostics: A step ahead

Background

Lung cancer diagnostics has witnessed a paradigm shift owing to rapid developments of agents for targeted molecular therapy. However, in a resource constrained country like India, panel based NGS cannot be made available to the population at large. Integration of artificial intelligence in cancer is showing promise in early detection, prognostication and therapeutic outcomes. After optimal successes in breast and esophageal cancers, lung cancer diagnostics, poses an unmet need for the same. NGS requires tedious workflows causing inadvertent delays in final report. Integrating a fairly accurate AI to predict oncogene addiction may help clinicians triage patients for NGS.

Methods

Whole slide images of hematoxylin and eosin (H&E)-stained slides of lung adenocarcinoma were used and split into multiple tiles and annotated for acinar, lepidic, micropapillary, papillary, or solid patterns. The tiles were divided into training, validation, and testing subsets. Training was conducted using VGG16 network (Keras Applications). The trained model was tested/validated with two independent datasets consisting of 5946 tiles each. In the next stage, the demographic and genomic data were combined with patterns to predict the mutational status of oncogenes such as EGFR, ALK, and ROS1. Multiclass ML algorithms such as random forest, decision tree classifier, extreme gradient boosting (XGBoost), GridSearchCV, and RandomizedSearchCV were implemented.

Results

3420 cases were used for training and 100 for prospective validation. The final model trained on VGG16 deep learning network (Keras Applications) performed at 87.22% validation and 80.49% testing accuracy conducted on two separate datasets(5946 image tiles each). Multiclass ML algorithms exhibited an accuracy of more than 90% for the mutation prediction (EGFR: 90.6%, ALK: 85.2%, ROS:79.3%). The number of cases in ALK and ROS1 positive groups were few, leading to lesser accuracy.

Conclusions

This initial experience suggests that integrating AI in lung cancer genomics is a value addition in the toolkit of the oncologist, aiding in prompt institution of appropriate therapy as well as triaging patients requiring upfront chemotherapy vs. panel based NGS testing.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Ullas Batra.

Funding

Conquer Cancer Foundation.

Disclosure

All authors have declared no conflicts of interest.