2020/10/19 by Kushal Mehta, Arshita Jain, Mehta, Kushal +9
Medicine · Computer Science · #Radiomics and Machine Learning in Medical Imaging #Lung Cancer Diagnosis and Treatment #AI in cancer detection
paper · pdf · doi:10.48550/arxiv.2010.11682
We present a hybrid algorithm to estimate lung nodule malignancy that\ncombines imaging biomarkers from Radiologist's annotation with image\nclassification of CT scans. Our algorithm employs a 3D Convolutional Neural\nNetwork (CNN) as well as a Random Forest in order to combine CT imagery with\nbiomarker annotation and volumetric radiomic features. We analyze and compare\nthe performance of the algorithm using only imagery, only biomarkers, combined\nimagery + biomarkers, combined imagery + volumetric radiomic features and\nfinally the combination of imagery + biomarkers + volumetric features in order\nto classify the suspicion level of nodule malignancy. The National Cancer\nInstitute (NCI) Lung Image Database Consortium (LIDC) IDRI dataset is used to\ntrain and evaluate the classification task. We show that the incorporation of\nsemi-supervised learning by means of K-Nearest-Neighbors (KNN) can increase the\navailable training sample size of the LIDC-IDRI thereby further improving the\naccuracy of malignancy estimation of most of the models tested although there\nis no significant improvement with the use of KNN semi-supervised learning if\nimage classification with CNNs and volumetric features are combined with\ndescriptive biomarkers. Unexpectedly, we also show that a model using image\nbiomarkers alone is more accurate than one that combines biomarkers with\nvolumetric radiomics, 3D CNNs, and semi-supervised learning. We discuss the\npossibility that this result may be influenced by cognitive bias in LIDC-IDRI\nbecause malignancy estimates were recorded by the same radiologist panel as\nbiomarkers, as well as future work to incorporate pathology information over a\nsubset of study participants.\n