Method of Applying Deep Learning Model (DLM) Trained on OCT Images of Lower Axial Resolution to Images of Higher Axial Resolution without Model Retraining for Retinal Layer Segmentation

Abstract

Purpose : We propose a method of applying DLMs trained on images obtained with a lower axial resolution OCT device to images obtained with higher axial resolution OCT devices without retraining the model.

Methods : A previously reported DLM (Wang & Birch, Front. Med. 2022), trained on the images obtained from standard Spectralis SD-OCT (axial resolution 3.87 μm/pixel), was used for automatic segmentation of retinal layers of OCT scan images obtained from two higher resolution devices: high-resolution Spectralis (HiRes-OCT, axial resolution 1.87 μm/pixel) and OHSU ultra-high resolution device (UHR-OCT, axial resolution 0.91 μm/pixel). Test images were 32 volume scans obtained from 12 patients with retinitis pigmentosa (26 HiRes-OCT scans from 8 patients and 6 UHR-OCT scans from 4 patients). Before applying the DLM model, all B-scan images in a volume scan were down-scaled vertically to match the axial resolution of SD-OCT. After the scaled images were segmented by the DLM, the segmentation map was rescaled back to the size of the original image. Manual correction (MC) was conducted by a human grader to check the accuracy of the rescaled segmentation for photoreceptor outer segment (OS) layer. Dice similarity and Bland-Altman analysis were conducted to assess the agreement between DLM-only and DLM-with-MC for the measurements of OS thickness, ellipsoid zone (EZ) area, and OS volume.

Results : Average dice coefficient ± SD between the EZ band segmentations determined by DLM and MC was 0.853 ± 0.137. For EZ area > 1 mm² (n=27), average dice coefficient ± SD between DLM and MC was 0.893 ± 0.097. Bland-Altman analysis revealed a mean difference ± SD of 3.060±2.293 μm, -0.945±2.011 mm², and -0.0019±0.0226 mm³, for OS thickness, EZ area and OS volume, respectively.

Conclusions : The preliminary results of dice score and OS metrics measurements were consistent with our previous findings, suggesting that the scaling-rescaling method would allow us to apply DLM trained on lower axial resolution images directly on OCT scan images obtained with higher axial resolution devices. With manual correction, accurate OS metrics measurements from higher resolution volume scans may be obtained. This method may provide a new tool to generate segmentation labels for higher axial resolution OCT images to be used in new model training and testing.

This abstract was presented at the 2024 ARVO Imaging in the Eye Conference, held in Seattle, WA, May 4, 2024.