Lung nodule malignancy classification using only radiologist-quantified image features as inputs to statistical learning algorithms: probing the Lung Image Database Consortium dataset with two statistical learning methods

Abstract

In the assessment of nodules in CT scans of the lungs, a number of image-derived features are diagnostically relevant. Currently, many of these features are defined only qualitatively, so they are difficult to quantify from first principles. Nevertheless, these features (through their qualitative definitions and interpretations thereof) are often quantified via a variety of mathematical methods for the purpose of computer-aided diagnosis (CAD). To determine the potential usefulness of quantified diagnostic image features as inputs to a CAD system, we investigate the predictive capability of statistical learning methods for classifying nodule malignancy. We utilize the Lung Image Database Consortium dataset and only employ the radiologist-assigned diagnostic feature values for the lung nodules therein, as well as our derived estimates of the diameter and volume of the nodules from the radiologists' annotations. We calculate theoretical upper bounds on the classification accuracy that are achievable by an ideal classifier that only uses the radiologist-assigned feature values, and we obtain an accuracy of 85.74 [Formula: see text], which is, on average, 4.43% below the theoretical maximum of 90.17%. The corresponding area-under-the-curve (AUC) score is 0.932 ([Formula: see text]), which increases to 0.949 ([Formula: see text]) when diameter and volume features are included and has an accuracy of 88.08 [Formula: see text]. Our results are comparable to those in the literature that use algorithmically derived image-based features, which supports our hypothesis that lung nodules can be classified as malignant or benign using only quantified, diagnostic image features, and indicates the competitiveness of this approach. We also analyze how the classification accuracy depends on specific features and feature subsets, and we rank the features according to their predictive power, statistically demonstrating the top four to be spiculation, lobulation, subtlety, and calcification.

Highly accurate model for prediction of lung nodule malignancy with CT scans

“…We treat the malignancy of nodules as a binary classification problem for “likely malignant” versus “likely benign” by thresholding the radiologist-assigned malignancy values so that malignancy values below 3 (i.e., 1 and 2) are categorized as benign and values above 3 (i.e., 4 and 5) are categorized as malignant. Recent published models for the problem of classifying S12 (considered as “likely benign”) versus the S45 (considered as “likely malignant”) include the models developed in 20 and 28 , 30 – 32 . In 20 , a multi-scale CNN (MCNN) approach is used to produce a feature vector of size 50 that is then input to a random forest classifier.…”

An interpretable deep hierarchical semantic convolutional neural network for lung nodule malignancy classification

“…By visual inspection of the ROC curves, HSCNN performs better than the traditional 3D CNN model. The area under the ROC curve (AUC) quantitatively compares the overall performance of a classification model and is frequently used as a metric to access performance in nodule classification [15,17,26,30,31]. Table 5 summarizes the mean AUC score, accuracy, sensitivity, and specificity for both models.…”

Standardized representation of the LIDC annotations using DICOM

“…The accompanied publications describing the dataset 4-6 accumulated over 1000 citations according to Google Scholar, and were used in a number of image analysis challenges 8 . Several tools (some of which have been released publicly) have also been contributed by the community to enable conversion of the XML annotations into alternative representations and to support exploration of the content [9][10][11][12] . Nevertheless, the XML annotations remain the only representation accessible to the users of the TCIA LIDCIDRI collection.…”

“…As an example, the same mechanisms for data encoding could be used for augmentation of the images and nodule annotations with the radiomics features derived from the nodule regions. This work utilizes tools developed earlier for interpreting XML annotations of LIDC 12 and for generating the standardized representations for image analysis results 14 . The dataset produced as a result of this work is harmonized with other standardized collections already in TCIA 15 .…”

“…First, this work leveraged the pylidc python package ( https://pylidc.github.io/ ) introduced by Hancock and Magnan 12 for accessing the volumereconstructed annotation contours for the individual scans and subjects, as extracted from the DICOM and XML components of the TCIA LIDCIDRI collection. pylidc provides an interface for iterating and querying various entities of the collection and their attributes.…”