PROJECT TITLE :

Automatic assessment of Pectus Excavatum severity from CT images using Deep Learning

ABSTRACT:

Pectus excavatum (PE) is the most common abnormality of the thoracic cage. The severity of PE is determined by extracting three indices from computed tomography (CT) images: Haller, correction, and asymmetry. This analysis has been carried out manually up until this point, which is a laborious process that is also prone to variability. In this paper, a fully automatic framework for PE severity quantification from CT images is proposed. The framework is comprised of three steps, which are as follows: (1) identification of the sternum's greatest depression point; (2) detection of eight anatomical keypoints relevant for severity assessment; and (3) measurements' geometric regularization and extraction. The first two steps are dependent on heatmap regression networks, which are built on the Unet++ architecture. These networks also use a novel variant that has been adapted to predict 1D confidence maps. On a database containing 269 CTs, the framework underwent an evaluation. Within a subset of patients, the intra-patient, intra-observer, and inter-observer variability of the estimated indices were analyzed for comparative purposes. The developed system demonstrated a good agreement with the manual method (with limits of agreement comparable to the inter-observer variability), with a mean relative absolute error of 4.41%, 5.22%, and 1.86% for the Haller, correction, and asymmetry indices, respectively. The proposed framework performed better than the expert in the intra-patient analysis, showing a higher reproducibility between indices extracted from different CT scans of the same patient. Overall, these results lend credence to the viability of the proposed methodology for the automatic, precise, and repeatable grading of PE severity within a clinical setting.