Your browser doesn't support javascript.
A Computational Technique Based on the Mean Curvature of Isophotes for the Detection and Classification of Interstitial Lung Diseases on Chest CT
American Journal of Respiratory and Critical Care Medicine ; 203(9), 2021.
Article in English | EMBASE | ID: covidwho-1277788
ABSTRACT
Rationale Thoracic computed tomography (CT) is an integral part of diagnosis and classification of interstitial lung diseases (ILD). However, recommended interpretative guidelines for ILD are often difficult to follow and radiological diagnosis of ILD frequently relies on subjective interpretation (gestalt) by the thoracic radiologist. Moreover distinguishing ILD from numerous other diagnoses that mimic the radiological appearance of ILD can at times be extremely challenging to non-specialized radiologists. Here we applied an image analysis technique of CT scans to explore automated discrimination between different forms of clinico-radiologically diagnosed ILD, various ILD mimickers and normal controls. This technique is based on the concept of mean curvature of isophotes (MCI), a computer vision technique recently shown to perform well at detecting COPD on CT scans. Here, we apply it for the first time to detection and classification of ILD.

Methods:

Our convenience sample of patients was divided in four categories 25 ILD patients with definite or probable usual interstitial pneumonia (UIP), 26 ILD patients indeterminate for UIP or with an ILD diagnosis other than UIP, 31 normal controls, and 36 ILD mimickers. The latter category included CTs featuring pulmonary edema, lymphangitic carcinomatosis, multifocal pneumonia, multifocal minimally invasive adenocarcinoma, pulmonary alveolar proteinosis, atelectasis, pulmonary sarcoidosis, pneumocystis jiroveci pneumonia, Kaposi sarcoma, and COVID19. From each patient's CT scan we computed MCI at a fine (1mm) and coarse (4mm) scale over a segmentation of the lungs performed with open-source software [http//chestimagingplatform.org]. We then quantified both MCI results with a probability density estimate (PDE). We subsequently performed a functional principal component analysis (FPCA) on the PDEs for each subject to capture the largest modes of variation. A multinomial logistic regression with the top 5 principal components from each scale as predictors was used to classify patient status.

Results:

Our method had excellent discriminative ability with a multi-class area under the operating characteristic curve of 0.976. The sensitivities and specificities ranged from 0.81-0.97 and 0.94-0.99, respectively (Table 1).

Conclusion:

The MCI technique showed good ability in differentiating definite and probable UIP from other categories of ILD, normal controls and various ILD mimickers. While the unsupervised nature of our predictors makes our method less susceptible to overfitting, external validation will be needed to provide an unbiased estimate of performance. The proposed method could ultimately be a valuable screening algorithm for incidental findings and/or decision support mechanism in contexts where specialized radiology expertise may not be readily available.

Full text: Available Collection: Databases of international organizations Database: EMBASE Language: English Journal: American Journal of Respiratory and Critical Care Medicine Year: 2021 Document Type: Article

Similar

MEDLINE

...
LILACS

LIS


Full text: Available Collection: Databases of international organizations Database: EMBASE Language: English Journal: American Journal of Respiratory and Critical Care Medicine Year: 2021 Document Type: Article