Virtual Library

Start Your Search

Yan Xu

Author of

  • +

    P2.11 - Screening and Early Detection (ID 178)

    • Event: WCLC 2019
    • Type: Poster Viewing in the Exhibit Hall
    • Track: Screening and Early Detection
    • Presentations: 1
    • Moderators:
    • Coordinates: 9/09/2019, 10:15 - 18:15, Exhibit Hall
    • +

      P2.11-30 - Effects of the Size of Nodules, Reconstruction Slice Thickness and Convolution Kernel on Radiomics Model in Classifying Pulmonary Nodules  (ID 2050)

      10:15 - 18:15  |  Presenting Author(s): Yan Xu

      • Abstract


      In recent years, the number of chest CT and LDCT scans for annual lung cancer screening has been increasing, the detection rate of the intermediate pulmonary nodules (IPNs) has increased, especially small pulmonary nodules (PNs). A non-invasive method be needed to early diagnosis the benign and malignant of IPNs, then it would be possible to reduce the false positives, missed diagnosis rate, and avoid overdiagnosis and over-treatment. The ability of radiomics to classify PNs by radiologists has been widely described, however, the detection performance of each radiomics varies greatly and the reproducibility was poor that are identified from these studies. Variability of acquisition parameters like contrast enhancement, slice thicknesses can affect the diagnostic performance of radiomic biomarkers. But there are few reports on the effect of PN size, reconstruction slice thickness and convolution kernels on the performance of radiomics in classifing PNs.


      We retrospectively collected 696 patients with 316 benign and 380 malignant PNs who underwent preoperative chest CT in the institution from March 1, 2015 to March 31, 2018. First, we analyzed the CT image of all the patients, and then we divided these images according to the nodule size and reconstruction kernel to test the diagnostic performance of the radiomic. 696 PNs were classified into three groups by nodule diameter: T1a (diameter ≤ 1.0 cm), T1b (1.0 cm < diameter ≤ 2.0 cm) and T1c (2.0 cm < diameter ≤ 3.0 cm). All CT images divided three groups according convolution kernels: Setting 1 (1mm/1.25 mm sharp), Setting2 (5 mm sharp), Setting 3 (5 mm smooth). Totally 1160 radiomic features were extracted from PNs segmentation on CT image delineated by an experienced radiologist. Sixteen radiomic models for predicting the malignancy of PNs in different size, reconstruction slice thickness and convolution kernels were built, respectively, based on the extracted radiomic features. Random selection of cases (70% Training and 30% testing) was employed to estimate the area under the receiver operating characteristic curve, accuracy, sensitivity and specificity to indicate the performance of the prediction models.


      The performance (AUC, accuracy, sensitivity and specificity) on prediction PN malignancy in different size PN with all convolution kernels were 0.817, 0.766, 0.807, 0.717 of all size-modal; 0.679, 0.756, 0.629, 0.796 of T1a-model; 0.700,0.690,0.757,0.594 of T1b-model and 0.703, 0.684, 0.673, 0.731 of T1c-model, respectively. AUC of all size PN with setting 1,2,3 group were 0.793, 0.800, 0.793, respectively. AUC was the highest in T1a with setting 2 model which equal 0.841, and the lowest in T1c with setting 4 which equal 0.625.


      Reconstruction slice thickness and convolution kernel have significant influence on the diagnosis performance of radiomics in classifying of less than 1cm PNs in CT images, and using 1 mm shin sharp reconstruction algorithm can obtain the best diagnosis performance in less than 1cm PNs. Big samples of PNs could alleviate the effect of reconstruction slice thickness and convolution kernel on radiomics in classifying of less than 3cm PNs in CT images and improve the diagnostic performance of radiomics of larger than 1cm PNs