Novel CADe/CADx System for Lung Nodules Segmentation and Classification on Computed Tomography Images
Vijayalaxmi Mekali1, Girijamma H. A2

1Vijayalaxmi Mekali, Department of Computer Science and Engineering, Kammavari Sangham Institute of Technology, Visvesvaraya Technological University, Bangalore (Karnataka), India.

2Dr. Girijamma H. A, Department of Computer Science and Engineering, R. N. S Institute of Technology, Visvesvaraya Technological University, Bangalore (Karnataka), India.

Manuscript received on 03 December 2019 | Revised Manuscript received on 11 December 2019 | Manuscript Published on 31 December 2019 | PP: 165-173 | Volume-9 Issue-2S December 2019 | Retrieval Number: B10941292S19/2019©BEIESP | DOI: 10.35940/ijitee.B1094.1292S19

Open Access | Editorial and Publishing Policies | Cite | Mendeley | Indexing and Abstracting
© The Authors. Blue Eyes Intelligence Engineering and Sciences Publication (BEIESP). This is an open-access article under the CC-BY-NC-ND license (

Abstract: Detection and classification of different types lung nodules poses major challenges in medical diagnosis routine. Classification of segmented nodules based on extracted hybrid features of segmented nodules have shown remarkable performance. Recently deep features alone and also with combination of hybrid features have improved nodules classification. In this research work new CADe/CADx system is proposed for detection and classification of Well Circumscribed Nodules, Juxta Vascular Nodules and Juxta Pleural Nodules. In nodules detection part, algorithms proposed in our previous work were used. Classifiers decision fusion based new nodules classification system is proposed. Four set of hybrid features and deep features using Convolution Neural Network are considered from segmented nodules. Hybrid features set consist of twenty four shape features, six GLCM features in four direction with a distance of two, six First Order Statistic features and twelve energy features. Five individually trained Probabilistic Neural Networks by all five set features separately used in nodule classification. In classification process all five classifiers decisions are fused at 2-level, 3-level, 4-level and 5-level. The proposed system achieved highest performance with 5-level fusion compared with other level fusions. System was evaluated on CT images of LIDC database with consideration of 2669 lung nodules of malignancy rate 1 to 5. Based on malignancy rate 2669 nodules are grouped as dataset 1 and dataset 2 with nodules of malignancy rate 1, 2, 3 and 3, 4,5 respectively. The 5-level decision fusion achieved highest accuracy of 95.72, sensitivity of 95.52, specificity of 95.79 and Area Under Curve of 96.21 for dataset 1 and accuracy of 92.54, sensitivity of 90.48, specificity of 94.63 and Area Under Curve of 92.69 for dataset 2.

Keywords: Computed Tomography, Computer Aided Detection/Diagnosis, Convolution Neural Network, Lung Cancer and Lung Nodule Classification.
Scope of the Article: Classification