Computer-aided detection of lung fibrosis remains a difficult task due to the small vascular structures, scars, and fibrotic
tissues that need to be identified and differentiated. In this paper, we present a texture-based computer-aided diagnosis
(CAD) system that automatically detects lung fibrosis. Our system uses high-resolution computed tomography (HRCT),
advanced texture analysis, and support vector machine (SVM) committees to automatically and accurately detect lung
fibrosis. Our CAD system follows a five-stage pipeline that is comprised of: segmentation, texture analysis, training,
classification, and display. Since the accuracy of the proposed texture-based CAD system depends on how precise we
can distinguish texture dissimilarities between normal and abnormal lungs, in this paper we have given special attention
to the texture block selection process. We present the effects that texture block size, data reduction techniques, and
image smoothing filters have within the overall classification results. Furthermore, a histogram-based technique to
refine the classification results inside texture blocks is presented.
The proposed texture-based CAD system to detect lung fibrosis has been trained with several normal and abnormal
HRCT studies and has been tested with the original training dataset as well as new HRCT studies. On average, when
using the suggested/default texture size and an optimized SVM committee system, a 90% accuracy has been observed
with the proposed texture-based CAD system to detect lung fibrosis.