Biomedical Image Processing | anirvandutta-bio

Pulmonary Cancer Cell Detection in Lung CT Images

Intern project, Computer Vision Lab, IIT Kharagpur

Project Guide - Dr. Sudipto Mukhopadhya, Electronics & Electrical Comm. Dept. , IIT Kharagpur

Project Members - Anirvan Dutta, Asish Dhara, Shrikant Mehre

This was the research project on biomedical image processing during Summer 2015. The research theme included biomedical image processing particularly the diagnosis of cancer cells in lung CT(computer tomography) images and developing a research-oriented approach to the existing technical challenges.

I worked on two major challenges -

Developing efﬁcient real time CAD system for pulmonary nodule detection in CT images.
Multi-feature analysis in classification of Pulmonary Nodules into Malignant and Benign

Research Topic - An efficient real time CAD system for pulmonary nodule detection in CT images.

Early detection of lung nodules is extremely important for the diagnosis and clinical management of lung cancer. In order for detection of cancer nodules we developed, a novel computer aided detection (CAD) system for the detection of pulmonary nodules in thoracic computed tomography (CT) imagery is presented. The report describes the architecture of the CAD system and assesses its performance on a publicly available database to serve as a benchmark for future research efforts. All the results were computed on the public database, Lung Image Database Consortium (LIDC). The CAD system uses a fully automated lung segmentation algorithm to define the boundaries of the lung regions. It combines intensity thresholding with morphological processing to detect and segment nodule candidates simultaneously.
A total of 120 geometric, intensity and gradient based feature has been proposed for efficient segmentation of cancer nodules from other lung nodules. We also present a 5-fold cross-validation performance analysis using Support Vector Machine was used to analyze the result on the LIDC collection.

Proposed Method.
The method is divided into four parts-

Lung Field Segmentation.
3D nodule candidate detection and segmentation.
3D nodule candidate feature extraction.
Classification using Support Vector Machine (SVM).

Lung Field Segmentation.
The first module of lung field segmentation included Local Contrast Enhancement Technique, to improve the details and local contrast of the lung CT image. Next, fixed threshold values in combination with some morphological and topological analysis are used to produce a preliminary segmentation of the lung region or mask. A threshold of 500 HU (Hounsfield Unit), which corresponds to a density between ‘water’ and ‘air’, is used to segment the body of the lung. A special operation of Contour Correction using chain code is used to correct the lung field. This ensured efficient detection of Juxta Pleural Nodules.

3D nodule candidate detection and segmentation
We use multiple gray level thresholding to segment the lung nodule candidates. However, here we pair each threshold operation with a specific morphological opening operation to produce a total of 15 intermediate candidate masks. Each intermediate candidate mask results from one threshold and opening operation. Further rule based analysis was done to remove non-nodule parts.

3D nodule candidate feature extraction.
After the candidates have been detected, a novel feature set comprising of 120 geometric, gradient and intensity features was used in feature extraction for classification. The features were both 2-D, computed on the biggest slice as well as 3-D, volume data.

SVM based classification.
SVM classifier was used on the basis of 120 features, to classify cancer nodule and lung nodule. A specificity of 0.95 and sensitivity of 0.88 was recorded on the basis of Receiver Operator Characteristic Curve.

Research Topic 2 - Multi-feature Analysis in classification of Pulmonary Nodules into Malignant and Benign.

Differentiation of malignant and benign pulmonary nodules is of paramount clinical importance. Various features reflect a powerful character of the malignancy like- texture, margin, geometric. The present report compares these features and presents a novel method of combining various features to increase the performance in classification. Quantitative comparison measures were made by the well-established Support Vector Machine classifier. The proposed feature set achieved an accuracy of 0.9515 in the area under the receiver operating characteristic curve which is more than the previous reported 0.9312 by Han et.al.
The segmentation of cancer nodules was semi-automatic unlike previous work on manual segmentation by radiologists. Three configuration was devised to report the result based on malignancy label.

Proposed Method
The entire work was divided mainly into two parts-

Semi-Automatic Segmentation of cancer nodule
Feature Extraction and Classification

Semi-Automatic Segmentation of cancer nodule
The boundary of the cancer nodule was segmented with help of seed point. This method involved gradient and texture analysis.

Feature Extraction and Classification
Various features comprising texture (Haralick), Histogram of Gradient, Margin based, Geometric were computed and each were analyzed. Final feature set comprising of 50 features was created and analyzed using SVM classifier.

Following were the classification result of the different feature set using SVM classifier. We got substantial improvement in detection and helped us publish two papers on this topic. Please check the Publication section.