Abstract
Purpose :
Optic nerve neuropathy, a spectrum of debilitating ophthalmic disorders, demands early, accurate diagnosis for effective intervention. Ultrasound (US), as a real-time, non-invasive, and cost-effective imaging modality, presents a promising avenue for optic nerve assessment. Our objective is to generate and curate datasets that can be used to train deep neural networks (DNNs) for inference on ocular and orbital US images, predicting (a) the correct diagnostic planes for estimating ocular biometric values, (b) structural orientation, and (c) anatomical features.
Methods :
We used point of care US (POCUS) to analyze tissue samples from optic nerve neuropathy patients and phantom tissue models. We evaluated image quality, diagnostic accuracy, optic nerve sheath diameter and dimensions of orbit in both phantom and real tissue images. We evaluated the performance of DNN models trained on US data generated using ocular phantom tissue. A convolutional neural network following the Network-in-Network architecture was trained to predict optic nerve sheath diameter, using the absolute error loss function. The training set consisted of 100 images with optic nerve diameters in the normal range and 40 images with optic nerve diameters in the abnormal range. The resulting models were then evaluated on real tissue samples not used during training, consisting of 100 normal images and 30 abnormal images.
Results :
We used 100 normal human eye scans as test subjects and 30 known abnormal phantom scans as training sets. The resulting training loss is 0.049, and the resulting testing loss is 0.15, indicating the model is overfit to the training set of phantom images. We also considered cases where the optic nerve sheath lies outside the normal range of 0.3 – 0.6 cm. On this, the trained model obtains a precision of 0.34, a recall of 0.83, and an F1 score of 0.49. This indicates that even the overfit model provides significant predictive skill in terms of detecting abnormal optic nerve sheath diameters and may have potential for the assessment of optic nerve head abnormalities in conditions like optic neuropathy.
Conclusions :
This study indicates the potential predictive skill when training DNNs from US images generated using ocular phantom tissue. The results show that even an overfit DNN has significant predictive skill on real tissue samples.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.