Abstract

Elevated intracranial pressure (ICP) is critical in patient outcomes for conditions like traumatic brain injuries, hydrocephalus, and intracranial hemorrhages. Effective ICP monitoring is essential for timely detection and intervention, reducing risks such as brain herniation and mortality. Conventional ICP monitoring methods, despite their accuracy, are invasive and can cause complications, including infection and brain tissue damage. Optic Nerve Sheath Ultrasonography (ONSUS) offers a promising non-invasive alternative by examining the optic nerve sheath (ONS) and its correlation with ICP. Elevated ICP causes the ONS to expand, measured as ONSD and correlates with increased ICP. ONSUS involves placing a high-frequency ultrasound probe on the closed eyelid, emitting sound waves that create an image of the ONS. The ONSUScan procedure starts with positioning the patient supine and slightly elevating the head. The healthcare provider then places the ultrasound probe on the closed eyelid to acquire the image.

US images can be noisy, with artifacts like speckle noise and shadowing. ONSUScan uses AI algorithms to address these challenges for image denoising and analysis. Convolutional Neural Networks (CNNs) are effective for feature extraction from noisy images using frameworks like TensorFlow. Autoencoders and U-Nets, such as DDUNet, enhance image quality by learning compressed representations and reconstructing denoised images. U-Nets use skip connections to preserve spatial information, improving denoising performance. Generative Adversarial Networks (GANs) further improve image quality by using a dual-network framework where the generator creates denoised images, and the discriminator enhances realism and accuracy. Wavelet-based denoising separates noise and signal components by transforming the image from the spatial domain to the wavelet domain. This process involves decomposing the image into multiple scales using wavelet transforms, denoising at each scale, and reconstructing the image from the modified wavelet coefficients. Open-source libraries like SciPy in Python offer robust wavelet transform functionalities, reducing speckle noise in US images. After denoising, the AI system estimates ICP by measuring ONSD using auto-segmentation techniques. A dataset of segmented ONSD images is prepared with tools like ITK-Snap, enabling automated segmentation in new images. Threshold-based classification identifies ONSD values indicating elevated ICP, typically above 5 mm in adults, corresponding to pressures exceeding 20 mmHg. Current technology measures ONSD with image enhancement algorithms in about 60 seconds, but new technologies are being developed with threshold-based auto-segmentation algorithms. 

dvanced regression models like Gradient Boosting Machines (GBMs) predict precise ICP values based on ONSD measurements and extracted image features. GBMs sequentially build an ensemble of models, with each new model correcting errors of the previous ones, improving accuracy. XGBoost, a widespread GBM implementation, is efficient and scalable, making it suitable for ICP prediction in ONSUScan. Open-source libraries like sci-kit-learn and XGBoost provide functionalities for building and training GBM models. ONSUScan is a novel technology because it leverages pre-existing, widely available ultrasound technology to provide a non-invasive and accurate method for monitoring intracranial pressure. This innovative approach integrates advanced AI algorithms for image denoising and analysis, enhancing the precision of optic nerve sheath diameter measurements. Its cost-effectiveness and non-invasiveness make it particularly suitable for resource-limited clinics in low- and middle-income countries, where traditional ICP monitoring methods may be inaccessible or impractical.