Machine Learning Ensemble Model for Improved Personalized Lung Cancer Risk Assessment and Malignant Nodule Detection

Overview

Objectives Current screening guidelines omit a large number of high-risk candidates that do not fit the traditional screening criteria. Furthermore, malignant lung nodule detection on CT scan is difficult as nodules are often miniscule and often benign/indeterminate. This causes radiologist screening of nodules to be expensive, low-throughput, and often inaccurate. This study develops an algorithm that utilizes machine learning and radiomics to build a complete lung cancer diagnostic pipeline that addresses these issues. Methods A large patient history dataset was obtained from Kaiser Electronic Medical Records and a separate large lung-CT scan dataset was compiled and hand-modified from various online sources. In order to assess a patient's risk of lung cancer, a 50-tree Gradient Boosted Machine (GBM) was constructed that employs personalized patient history variables including age, prescriptions, ethnicity, body mass index, blood pressure, and diagnoses to better assess true risk of patients. Once a CT scan is conducted to identify malignant lung nodules, an ensemble of 3D Convolutional Neural Networks (CNNs) of discriminator VGG- like and U-net architectures, trained with multitudinous augmentations and gradient clipping on a hand- engineered dataset, determines nodule morphology (calcification, spiculation, size), position, and malignancy. From these features, a linear classifier predicts lung cancer development in one year. Results The GBM significantly surpasses current guideline assessments, capturing omitted patient groups at high risk for lung cancer (sensitivity increased from 23% to 88%). Moreover, the CNN Ensemble obtained statistically comparable predictions to radiologist readings of scans. Further, the Ensemble substantially reduced the false positive rate of Computer Aided Diagnosis models (from on average 15.28 to on average 1.68 false positives per scan). Conclusions This model could serve as primary screen for lung cancer nodules to decrease radiologist involvement in screening. The combined automated lung cancer diagnostic system increases early-detection rates and reduces false positive rates of Computer Aided Diagnosis systems, thereby greatly improving the timeliness, accuracy, and affordability of lung cancer detection.

Awards (1)

• Second Place

Competition history

• CSEF 2019 Computational Systems & Analysis (Senior Division Only) · Entry S0805

Resources

• CSEF project summary ↗

Related projects

Closest projects by meaning, across every fair and year in the corpus.

Browse more like this

• All Systems Software and Computer Science projects
• CSEF Systems Software and Computer Science projects
• Every CSEF 2019 project
• CSEF Second Place winners