Semi-automatic classification of textures in thoracic CT scans. (20th July 2016)

• Record Type: Journal Article
• Authors: Kockelkorn, Thessa T J P; de Jong, Pim A; Schaefer-Prokop, Cornelia M; Wittenberg, Rianne; Tiehuis, Audrey M; Gietema, Hester A; Grutters, Jan C; Viergever, Max A; van Ginneken, Bram
• Is Part Of: Physics in medicine & biology. Volume 61:Number 16(2016:Aug.)
• Page Start: 5906
• Page End: 5924
• Publication Date: 2016-07-20
• Subjects: interactive classification -- texture analysis -- supervised learning -- interstitial lung disease; Biophysics -- Periodicals; Medical physics -- Periodicals; 610.153
• DOI: 10.1088/0031-9155/61/16/5906
• Languages: English
• ISSNs: 0031-9155

Abstract

Abstract: The textural patterns in the lung parenchyma, as visible on computed tomography (CT) scans, are essential to make a correct diagnosis in interstitial lung disease. We developed one automatic and two interactive protocols for classification of normal and seven types of abnormal lung textures. Lungs were segmented and subdivided into volumes of interest (VOIs) with homogeneous texture using a clustering approach. In the automatic protocol, VOIs were classified automatically by an extra-trees classifier that was trained using annotations of VOIs from other CT scans. In the interactive protocols, an observer iteratively trained an extra-trees classifier to distinguish the different textures, by correcting mistakes the classifier makes in a slice-by-slice manner. The difference between the two interactive methods was whether or not training data from previously annotated scans was used in classification of the first slice. The protocols were compared in terms of the percentages of VOIs that observers needed to relabel. Validation experiments were carried out using software that simulated observer behavior. In the automatic classification protocol, observers needed to relabel on average 58% of the VOIs. During interactive annotation without the use of previous training data, the average percentage of relabeled VOIs decreased from 64% for the first slice to 13% for the second half of the scan. Overall, 21% of the VOIs were relabeled. When previous training data was available, the average overall percentage of VOIs requiring relabeling was 20%, decreasing from 56% in the first slice to 13% in the second half of the scan.