Online Poster Portal

  • Author
    Abhinav Suri
  • Discovery PI

    William Hsu PhD

  • Project Co-Author

    Ashley Prosper MD

  • Abstract Title

    Automated Detection and Classification of Indeterminate Pulmonary Nodules on CT using Deep Learning

  • Discovery AOC Petal or Dual Degree Program

    Informatics & Data Science

  • Abstract

    Specialty (if any): Radiology

    Keywords: Artificial Intelligence, Lung Cancer, Screening

    Background:

    Lung cancer affects thousands of Americans each year; however, screening for cancer in at-risk populations using computed tomography (CT) scans can lead to early detection of cancer. While imaging features of detected nodules inform the likelihood of malignant transformation, progression of nodules in the 6-30 mm range are “indeterminate”, requiring frequent follow up scans. The study of how these nodules change over time is hindered by the lack of automated tools that can detect and characterize their consistency (which is thought to correlate with malignant potential).

    Objective: We sought to create a tool that could detect and characterize the consistency of indeterminate pulmonary nodules in an automated manner.

    Methods: Lung nodules across three datasets (National Lung Screening Trial [NLST], Lung Image Database Consortium [LIDC], and an internal dataset from [UCLA]) were annotated by radiologists. Four AI models were trained to detect lung nodules. 80% of the NLST and LIDC datasets were used for training models; the remainder of those datasets and the UCLA dataset were used for testing. Detected nodules from the best algorithm were used to create a neural network (DenseNet121) to classify nodules as solid, part-solid, non-solid, or cystic. Detection rates and area under the receiver operator characteristic curve (AUC) are reported on the test set.

    Results: The final dataset had 2670 patients and 5658 lung nodules (822 patients, 1361 nodules in test set). Our final algorithm detected 95% of nodules on the test set. The nodule classification model achieved an overall AUC of 0.8334 (best at detecting non-solid nodules with an AUC of 0.8850, worst for cystic nodules: 0.8048).

    Conclusions: We were able to create a tool to fully automate detection and classification of indeterminate pulmonary nodules in a multi-center population. Future studies include using this algorithm in a prospective cohort to characterize the natural history of indeterminate pulmonary nodules.