Study Purpose Method Database Observations Kawata et al. [227 ] To characterize morphology of small pulmonary nodules Using surface curvatures and a ridge line Thin-section CT images for 56 cases including 42 M and 14 B nodules The distribution of the nodule characteristics in the feature space shows good evidence of separation between the two classes Henschke et al. [230 ] To explore the usefulness of neural networks (NNs) to help in this differentiation Statistical-multiple-object detection and location system (S-MODALS) NNs technique developed for automatic target recognition (ATR) CT images of 28 pulmonary nodules, 14 B and 14 M , each having a diameter less than 3 cm were selected Correctly identify all but three B nodules, but did not misclassify any M nodule Kawata et al. [137 ] To characterize the internal structure of small pulmonary nodules Using multiscale curvature-based shape spectrum Thin-section CT images of 27 pulmonary nodules B and 18 solid and infiltrative M cases) The distribution of the nodule characteristics in the feature space shows good evidence of separation between the two classes McNitt-Gray et al. [231 ] To classify nodules into benign or malignant LDA with stepwise feature selection based on nodule's shape, size, attenuation, distribution of attenuation, and texture HRCT scans of 17 M and 14 B nodules Correct classification rate of 90.3% Kawata et al. [232 ] To discriminate between B and M nodules LDA with stepwise feature selection based on nodule's features (density and curvatures) and surrounding structure features CT images of 248 pulmonary nodules including 179 M and 69 B nodules Nodule's features (
= 0.88) were more effective than the surrounding structure features (
= 0.69) in classification. Combing both features achieves
= 0.94 Kawata et al. [233 ] To obtain nodule diagnosis information by image retrieval from a database of known diagnosis Retrieving the nodules with similar characteristics from a 3D image database based on its CT density and curvature index CT images of 248 pulmonary nodules including 179 M and 69 B nodules The resulted visual figures are sorted from more similar to less similar with M case and show a high similarity with the test nodule Matsuki et al. [234 ] To classify nodules into benign or malignant ANN with 16 subjective features determined by radiologists and 7 clinical data 155 HRCT scans of 99 M and 56 B nodules
= 0.951 in a leave-one-out test Lo et al. [235 ] To quantify lung nodules in thoracic CT A NNs based on geometrical features, intensity, and texture features CT images of 48 cases of lung nodules (24 B , 24 M )
= 0.89 Aoyama et al. [236 ] To classify nodules into benign or malignant LDA with Wilks' lambda stepwise feature selection Thick-slice (10 mm) screening LDCT scans of M and 413 B nodules
= 0.846 in a leave-one-out test Nakamura al. [237 ] To classify nodules into benign or malignant Two NNs: one trained with 8 subjective features recorded by radiologist rating and the other with 12 matched computerized objective features 56 radiographs of 34 M and 22 B nodules
= 0.854 using subjective features and
= 0.761 using objective features. The reported radiologist accuracy was
= 0.752 Iwano et al. [238 ] To classify the shape of pulmonary nodules using computer analysis of HRCT LDA with 2 features (circularity and second moment) HRCT images from 102 patients with 102 nodules classified as round or oval, lobulated, polygonal, tentacular, speculated, ragged, and irregular For 95 of 102 cases, the shape classification by the two radiologists was the same. For the seven mismatched cases, pulmonary nodules with circularity ≤0.75 and second moment
0.18 were very likely to reveal lung cancer Shah et al. [239 ] To classify nodules into benign or malignant Logistic regression or QDA with stepwise feature selection from 31 features Thin-slice (≤3 mm) CE-CT scans of 19 M and 16 B nodules
= 0.69 and 0.92 with logistic regression and QDA, respectively, in a leave-one-out test Matsuoka et al. [240 ] To analyze features of peripheral noncalcified solitary pulmonary nodules in patients with emphysema Analyze the fractal dimensions of the nodule interfaces, nodule circularity, and the percentage of the nodule surrounded by emphysema CT images of 41 nodules (21 M , 20 B ) in 41 patients with emphysema In patients with emphysema, there were no significant differences in fractal dimension, circularity, or frequency of lobulation or spiculation between M and B nodules Mori et al. [241 ] To classify nodules into benign or malignant LDA using 3 features: shape index, curvedness values, and attenuation Thin-slice (2 mm) CE-CT scans of 35 M and 27 B nodules
= 0.91 and 1.0 with non-CE CT and CE-CT, respectively, in a leave-one-out test Suzuki et al. [117 ] To classify nodules into Benign or Malignant Multiple MTANNs using pixel values in a
subregion Thick-slice (10 mm) screening LDCT scans of M and 413 B nodules
= 0.88 in a leave-one-out test Iwano et al. [242 ] To classify nodules into benign or malignant LDA based on nodule's circularity and second moment features HRCT (0.5–1 mm slice) scans of 52 M and 55 B nodules Sensitivity of 76.9% and a specificity of 80% Way et al. [243 ] To classify nodules into benign or malignant LDA or SVM with stepwise feature selection CT scans of 124 M and B nodules in 152 patients
= 0.857 in a leave-one-out test Chen et al. [244 ] To classify nodules into benign or malignant ANN ensemble CT scans (slice thickness of 2.5 or 5 mm) of 19 M and 13 B nodules
= 0.915 in a leave-one-out test Lee et al. [245 ] To classify nodules into benign or malignant GA-based feature selection and a random subspace method Thick-slice (5 mm) CT scans of 62 M and 63 B nodules
= 0.889 in a leave-one-out test El-Baz et al. [246 ] To classify nodules into benign or malignant Analysis of the spatial distribution of the nodule Hounsfield values CT scans (2 mm slice) of M and 58 B nodules Sensitivity of 92.3% and a specificity of 96.6% El-Baz et al. [247 ] To classify nodules into benign or malignant Analysis of the SHs needed to delineate the lung nodule CT scans (2 mm slice) of 153 M and 174 B nodules
= 0.9782
