Automation of ultrasonographic optic nerve sheath diameter measurement using convolutional neural networks

• DOI: 10.1111/jon.13163
• Publication Date: 2023-10-17T04:23:52Z
• Authors: Mohammad I. Hirza...; Supratik Bose; Jingtong Hu; Jonathan S. Maltz
• Keywords: Mean absolute error

Abstract Background and purpose Ultrasonographic optic nerve sheath (ONS) diameter is a noninvasive intracranial pressure (ICP) surrogate. ICP monitored invasively in specialized intensive care units. Noninvasive monitoring important less settings. However, using ONS (ONSD) limited by the need for experts to obtain perform measurements. We aim automate ONSD measurements deep convolutional neural network (CNN) with novel masking technique. Methods trained CNN reproduce masks that mark ONS. The edges of mask are defined an expert. Eight models were 1000 epochs per model. Dice‐similarity‐coefficient‐weighted averaged outputs eight yielded final predicted mask. hundred seventy‐three images obtained from 52 transorbital cine‐ultrasonography sessions, performed on 46 patients brain injuries. fourteen 48 scanning sessions used training validation 59 four testing. Bland‐Altman Pearson linear correlation analyses evaluate agreement between expert Results Expert CNN‐derived estimates had strong ( r = 0.7, p < .0001). mean (standard deviation) 5.27 mm (0.43) compared estimate 5.46 (0.37). Mean difference (95% confidence interval, value) 0.19 (0.10‐0.27 mm, .0011), root square error 0.27 mm. Conclusion A can learn measurement without image segmentation or landmark detection.