Abstract
Purpose :
Retinoblastoma (RB) is the predominant pediatric intraocular tumor and is manually detected and evaluated from Optical Coherence Tomography (OCT) images of the patient’s retina to inform treatment planning. The evaluation of RB from OCT requires the clinician to track tumor morphology and treatment effects over multiple patient visits which is time-consuming and subject to human error. Therefore, we addressed this challenge by developing a machine learning model that can segment out RB and scar regions from retrospective OCT images of a patient.
Methods :
We analyzed one RB patient’s OCT images collected over 7 visits from 2021 to 2022. A total of 888 images of 1000x1024 pixels were manually annotated to segment out RB and the scars formed by treatment. Images were then resized to 512x512. 80% of the annotated dataset was used to train a machine learning model and 20% to test the performance of the model. The model was evaluated by comparing regions it segmented with the annotations of the test dataset using the Jaccard score metric to quantify overlap. When the model segmented incorrect pixels for a class, we measured their pixel distance to the nearest manual annotation to evaluate proximity to ground truth.
Results :
Our model achieved Jaccard scores of 0.75 and 0.59 for the RB and scar respectively. Literature has shown that OCT segmentation models can achieve Jaccard scores greater than 0.8 for segmenting retinal layers. However, our OCT images were noisy as they had not been averaged during acquisition. Also, the segmentation of scar was challenging due to its irregular shapes, small size and lack of sufficient images available. When examining the proximity of the model’s segmentations to the annotations, it was observed that non-overlapping pixels were an average of 13 ± 19 pixels away for RB and 9 ±15 pixels away for scar in an image size of 512x512. Qualitative observations of the segmented images also indicated that the model’s segmentations were close to the annotations when there was not a perfect overlap.
Conclusions :
The model was able to segment out necessary regions for detecting RB and scar from OCT images. In the cases where it did not perfectly segment out the regions, its proximity was measured relative to the ground truth. Examining the model’s performance on more patient data will refine the model and alleviate the current manual approach of detecting the regions for RB treatment planning.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.