Automatic Size Measurement of Cavities on Chest Radiographs Using Supervised Learning and Dynamic Programming

P. Maduskar, L. Hogeweg, H. Ayles, R. Dawson, P. de Jong and B. van Ginneken

Annual Meeting of the Radiological Society of North America 2011.

PURPOSE Accurate measurement of the size of cavities on chest radiographs (CXRs) is important for tuberculosis (TB) treatment monitoring and to make the decision to start TB treatment in the case of smear-negativeTB suspects. An automatic technique for cavity segmentation is presented and compared to inter-reader agreement of human experts. METHOD AND MATERIALS A data set of 105 digital CXRs (2048 A-A?A 1/2 2048 resolution, pixel size 0.25 mm, Delft Imaging Systems, The Netherlands) was collected at Kanyama Health Clinic, Lusaka, Zambia and University of Cape Town Lung Institute, Cape Town, South Africa. For training and system development, 20 CXRs with one cavity and 30 normal cases were used. For evaluation, 55 CXRs with one cavity were used. These cavities were manually outlined by three human experts, one chest radiologist and two readers certi??ed to read CXRs according to the Chest Radiograph Reading and Recording System. Cavities are automatically segmented after the user clicks near the cavity center. A pixel classi??er was trained to construct a cavity wall likelihood map. The classifier uses Gaussian, location and Hessian features and was trained with pixels within 1 mm of a cavity border as positive examples and random pixels from normal CXRs as negative examples. A polar transformation of the likelihood map around the center point was used as a cost function to search for an optimal border using dynamic programming. Cavity segmentations are compared using Jaccard overlap measure. Cases where the average overlap between pairs of manual expert segmentations is above/below 0.80 are defined as obvious/challenging cases. RESULTS The evaluation data contained 37 obvious and 18 challenging cases. The average overlap between manual expert segmentations was 0.87 A-A?A 1/2 0.02 and 0.67 A-A?A 1/2 0.05 for the obvious and challenging cases, respectively. For automatic versus manual segmentation the average overlap was 0.77 A-A?A 1/2 0.01 and 0.66 A-A?A 1/2 0.01, respectively. CONCLUSION Cavity segmentation is a challenging task with considerable disagreement between human expert readers. Our automated algorithm shows results comparable to experts and can be considered as an effectual and consistent method for cavity size measurements in CXR. CLINICAL RELEVANCE/APPLICATION Automatic cavity segmentation on chest radiographs can facilitate TB treatment follow-up and help to make a clinical decision regarding the start of TB treatment.