Utilização de visão computacional para a reconstrução automatizada da área de secção transversa muscular a partir de imagens sequenciais obtidas por ultrassonografia

Abstract

The aim of the present study was to propose and validate a tool based on computer vision techniques that allows for the automated reconstruction (AR) of the vastus lateralis (VL) muscular cross-sectional area (MCSA) from sequential images obtained using an ultrasound (US) machine. Methods: Four hundred and eighty-eight VL US image sequences were used for VL MCSA reconstruction. Two different reconstruction techniques were utilized. For the already validated manual reconstruction (MR) technique, the sequential images were manually adjusted until the MCSA of the VL was fully visible for each image sequence. For AR, computer vision techniques were combined in a tool capable of automatically reconstructing the MCSA of the VL based on the steps described for the proper application of MR. After the quantification in cm² of all VL MCSA by both MR (n = 488) and AR (n = 488) techniques, the results were used to investigate the validity of the AR in measuring VL MCSA from of sequential images of the VL obtained by US. Our findings demonstrated good validity with low coefficient of variation values (1.51%) for AR compared to MR. The Bland-Altman plot showed low bias (-0.01 cm²; IC95% = 0.04, -0.06) and close limits of agreement (+1.18 cm², -1.19 cm²) containing 95% of the comparisons. Conclusion: The AR technique is valid compared to MR when measuring VL MCSA in a heterogeneous sample.