Awais Mansoor PhD, Presenter: Nothing to Disclose

Ulas Bagci PhD, MSc, Abstract Co-Author: Nothing to Disclose

Brent Foster, Abstract Co-Author: Nothing to Disclose

Ziyue Xu PhD, Abstract Co-Author: Nothing to Disclose

Jayaram K. Udupa PhD, Abstract Co-Author: Nothing to Disclose

Daniel Joseph Mollura MD, Abstract Co-Author: Nothing to Disclose

To develop a fast and fully automated image segmentation method that automatically delineates the pathological lungs from computed tomography (CT) scans and provides precise quantitative analysis.

With IRB approval, 200 CTs (with varying pathology levels) and 40 CT scans (healthy controls) were collected and used for the experiments. The pathological lungs were divided into 4 categories based on the pathological volume measured by participating expert radiologists during reference standard construction—control, minimum pathology, medium pathology, and heavy pathology. Our algorithm is based on the fuzzy connectedness image segmentation method integrated into a patient specific shape model. By this model creation, we identified the possible regions where segmentation may fail. For those regions, we refine our segmentation through a random forest machine learning algorithm by accurately labeling tissue types as normal or abnormal within the lungs. The flow diagram of the proposed method is provided in Fig. 1.

An average Dice Similarity Coefficient of >90% was obtained—100% indicates the reference standard provided by the expert radiologists. A user interface was designed for radiologists to use the proposed quantification and evaluation system in their daily tasks. To the best of our knowledge, this is the first fully automated, robust, and accurate method using patient specific shape model within the pathological lung segmentation literature.

We developed a novel fully automated segmentation technique for lung segmentation, the initial analysis show promising scope for the technique in accurately segmenting pathological and normal lungs in routine clinical environment.

Our proposed method is an automated system with applications that can be used in clinics and provide accurate quantification of lung diseases for early prognosis or pre-op assessments.

Mansoor, A, Bagci, U, Foster, B, Xu, Z, Udupa, J, Mollura, D, A Robust Pathological Lung Segmentation Platform Using Fuzzy—Connectedness with Patient-specific Modeling.  Radiological Society of North America 2013 Scientific Assembly and Annual Meeting, December 1 - December 6, 2013 ,Chicago IL. http://archive.rsna.org/2013/13044303.html