Abstract
Interstitial-lung-disease progression assessment and diagnosis via radiological findings on computed tomography images require significant time and effort from expert physicians. Accurate results from these analyses are critical for treatment decisions. Automatic semantic segmentation of radiological findings has been developed recently using convolutional neural networks (CNN). However, on the one hand, few works present individual performance scores for radiological findings that allow for accurately measuring fibrosis segmentation performances; on the other hand, the poorly annotated quality of available databases may mislead researcher observations. This study presents a CNN methodology employing three different architectures (U-net, LinkNet, and FPN) with transfer learning and data augmentation to enhance the performance in semantic segmentation of fibrosis-related radiological findings (FRF). In addition, considering the poor quality of manual CT tagging on available datasets, we use two alternative evaluation strategies, first using only the fibrosis region of interest. Second, re-tagging and validating the test set by an expert pulmonologist. Using DICOM images from the Interstitial Lung Diseases Database, the implemented approach achieves a Jaccard Score Index of 0.7355 with a standard deviation of 0.0699 and a Dice Similarity Coefficient of 0.8459 with a standard deviation of 0.0470 comparable to state-of-the-art performance in FRF semantic segmentation. Also, a visual evaluation of the images automatically tagged by our proposal was performed by a pulmonologist. Our proposed method successfully identifies these FRF areas, demonstrating its effectiveness. Also, the pulmonologist revealed discrepancies in the dataset tags, indicating deficiencies in FRF annotations.