RSNA 2017

Abstract Archives of the RSNA, 2017


Deep Learning-Based Content-Based Image Retrieval for Finding HRCT Images of Similar Patients with Interstitial Lung Disease: Validation with 100 Paired HRCTs and Automatic Quantification of Six Disease Patterns

Kyu-Hwan Jung, PhD, Seoul, Korea, Republic Of (Presenter) Stockholder, VUNO Inc.
Ilji Choi, MS, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Sangkeun Kim, Seoul, Korea, Republic Of (Abstract Co-Author) Employee, VUNO Inc
Namkug Kim, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Stockholder, Coreline Soft, Inc
Beomhee Park, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Eunsol Lee, MD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Sangmin Lee, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose
Joon Beom Seo, MD, PhD, Seoul, Korea, Republic Of (Abstract Co-Author) Nothing to Disclose


Interstitial lung diseases (ILDs) represent a major cause of morbidity and mortality. High-resolution computed tomography (HRCT) has become critical to characterize the imaging patterns of ILD, but this approach remains vulnerable to inter- and intra-observer variation. To overcome human variation, automated techniques have been applied for differentiating a variety of obstructive lung diseases based on the features of a density histogram and texture analyses. Quantitative assessment of lung parenchymal texture is important to analyze and differentiate regional diseased patterns of ILD, which would lead to content-based image retrieval (CBIR). Using deep learning technique with Siamese convolutional neural net (S-CNN) on the raw image itself and classified disease patterns, 3D CBIR at HRCT is potentially useful for diagnosis and decision by retrieval of similar HRCT to referring similar patient previously diagnosed with known treatment response and survival. To address these unmet clinical needs, we have developed DILD CBIR platform, a deep learning-based CBIR system and its evaluation tool with known 100 paired HRCTs of the same patient; thus to provide an efficient and reliable quantification for the assessment of CBIR performance for ILD patients.


The whole CBIR system in this demonstration can be decomposed into three steps: 1)Segmentation of lung parenchyma area in HRCT, 2)Quantification of a lung in terms of ILD, 3)Extraction of representative features from quantified HRCT and retrieval of most similar cases in terms of extracted features. For the lung parenchyma segmentation, we used semantic segmentation approach using CNN with auxiliary spatial information to reduce false positives in the lower lobe. By training the network with 349 diseased HRCT cases (134 IPF, 128 COP, and 87 NSIP) and expert-labelled lung masks, we could obtain dice similarity coefficient of 0.9872. For the ILD quantification using HRCT images, six local disease patterns of ILD (normal, ground-glass opacity, reticular opacity, honeycombing, emphysema, and consolidation) were determined. A total of typical 900 regions of interest were marked by an experienced radiologist and used for training CNN. The trained model with test classification accuracy of 96.72% was used to label the whole lung area using fully-convolutional approach. Each ILD case is then quantified by characterizing the distribution of the disease patterns: area fraction of each pattern, directional probabilistic density function, and regional cluster distribution pattern. Finally, we trained S-CNN to learn similarity metric between two quantified HRCTs. Using same HRCT cases for training lung parenchyma segmentation(349 cases) the network was trained to discriminate if given pair of HRCT cases is from same disease class or not. The trained model provides feature representation of each HRCT in the space of ILD cases which can be used to retrieve similar cases based on the distance in this space. To evaluate our CBIR system, additional 100 pairs of initial and follow-up (within 2 years) HRCT images are collected. The radiologist reviewed all pairs and rated them using 3 similarity labels (1: exactly same (51%), 2: similar (32%), 3: progress observed (17%)). We measured how the method accurately finds the same pair having score1 within k nearest neighbors in the trained metric space. As shown in Table 1 (b), our method has 79.17% and 77.78% recall accuracies when 5 neighbors are considered. We have shown the feasibility of the CBIR for ILD using lung quantification and metric learning. The result shows the proposed method can retrieve similar cases significantly. This CBIR method for finding a similar patient of ILD can be used for clinical diagnosis and differential diagnosis on DILD with referring previous similar cases.


The purpose of this demonstration is to showcase a DILD CBIR platform for retrieval of similar HRCT cases in the HRCT repository with an easy web interface. The educational demonstration of DILD CBIR platform will use computer-based hands-on demonstration at RSNA. We will set up a cloud platform of DILD CBIR platform with use of multiple computers, one for the thin client for CBIR and the other for the demonstration of automatic quantification and feature visualization of HRCT. The demonstration will cover the entire workflow ranging from image acquisition protocol, automated post-processing, interactive reviewing, automated CBIR, advanced analysis and structured reporting, and will select patient cases from our clinical study approved by institutional review board of Asan Medical Center, which has been anonymized in accordance with the HIPAA Privacy Rule.


Medical Physics (3.761), 2013 May;40(5):051912 Journal of Digital Imaging(1.421), 2011 Dec;24(6):1133-40, 21311944 arXiv,:1606.00915, 2016 Jun arXiv:1605.06211, 2016 May

Meet-the-Experts Schedule:

Monday 12:15pm - 1:15pm Tuesday 12:15pm - 1:15pm Wednesday 12:15pm - 1:15pm