Toward Understanding the Size Dependence of Shape Features for Predicting Spiculation in Lung Nodules for Computer-Aided Diagnosis

Abstract

We analyze the importance of shape features for predicting spiculation ratings assigned by radiologists to lung nodules in computed tomography (CT) scans. Using the Lung Image Database Consortium (LIDC) data and classification models based on decision trees, we demonstrate that the importance of several shape features increases disproportionately relative to other image features with increasing size of the nodule. Our shaped-based classification results show an area under the receiver operating characteristic (ROC) curve of 0.65 when classifying spiculation for small nodules and an area of 0.91 for large nodules, resulting in a 26 % difference in classification performance using shape features. An analysis of the results illustrates that this change in performance is driven by features that measure boundary complexity, which perform well for large nodules but perform relatively poorly and do no better than other features for small nodules. For large nodules, the roughness of the segmented boundary maps well to the semantic concept of spiculation. For small nodules, measuring directly the complexity of hard segmentations does not yield good results for predicting spiculation due to limits imposed by spatial resolution and the uncertainty in boundary location. Therefore, a wider range of features, including shape, texture, and intensity features, are needed to predict spiculation ratings for small nodules. A further implication is that the efficacy of shape features for a particular classifier used to create computer-aided diagnosis systems depends on the distribution of nodule sizes in the training and testing sets, which may not be consistent across different research studies.

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

“…They achieve a sensitivity of 90%. Niehaus et al 11 analyzed the association between nodule size and the success of computing accurate spiculation labels from the LIDC dataset. They find that the AUC score increases roughly from 0.6 to 0.9 as the size of the nodules considered increases.…”

“…However, we observe that there are a number of published articles that employ these physician-quantified labelings of spiculation and lobulation from the LIDC dataset, but none mention the possible mislabelings in the dataset nor the exclusion of these 399 cases from their studies. [6][7][8]10,11,18,22,32,[38][39][40][41] Leaving out these 399 cases, we are left with 4384 nodule annotations that were consistently labeled. The number of annotations used is further reduced from 4384 to 2817 to exclude indeterminate cases (as described in Sec.…”

“…Nevertheless, many studies have quantified such features numerically for the purpose of either computer-aided diagnosis (CAD) or computer-aided characterization, by mathematically approximating characteristics of the features (from an interpretation of their respective qualitative definitions) using an assortment of algorithmic methods. [5][6][7][8][9][10][11][12] On the other hand, others have used the algorithmic quantification of image features only as intermediate quantities within a system for classifying nodules as malignant or benign. [13][14][15][16][17][18] In these approaches, it is not clear how well the quantified features capture the true physical nature of the features themselves because only error metrics for the accuracy of nodule classification are considered rather than the approximation error in quantifying the features themselves.…”

Visual-Acoustic Sensor-Aided Sorting Efficiency Optimization of Automotive Shredder Polymer Residues Using Circularity Determination

“…The border line thus obtained is a combination of lines between the centers of two adjacent border pixels, as the red line shows in Figure 7b. The perimeter is calculated according to the following equation [38]: P=true∑j=1n−1D(bj,bj+1) where b j and b j +1 are the adjacent border pixels and D equals 0.948 when b j and b j +1 are on a straight line; otherwise, D equals 1.343, which has been confirmed by Niehaus [41].…”

Ensemble classification for predicting the malignancy level of pulmonary nodules on chest computed tomography images

“…When making a diagnosis using medical imaging, physicians rate the characteristics (texture, margin, lobulation and calcification) of pulmonary nodules using empirical and subjective methods to determine their malignant phenotype (3)(4)(5)(6)(7). This method is subjective and highly dependent on the physician's experience.…”