Multiple pathology detection in macular OCT B-scans using localized-3D information

Abstract

Purpose : The advent of data-intensive imaging technologies, along with the increasing number of patients, creates a need for an automatic triaging tool for improving workflow efficiency. In this study, we developed a multiple-pathology detector that can be used as a clinical decision support tool for macular OCT cubes.

Methods : This retrospective study used 76,544 B-scans from macular cubes (512x128) of 598 subjects (250 normals, 348 with ocular pathologies) acquired using CIRRUS™ HD-OCT 5000 (ZEISS, Dublin, CA). Each B-scan is labeled by the categories listed in figure 1 or as ungradable by two retina specialists. If there were a difference in the cube level grading, the ground truth was finalized after an adjudication by a third specialist and the B-scan level labels were created by combining the gradings of both specialists using ‘OR’ logic.

A 5-fold cross-validation with 80%-20% split was used. The training set was augmented to create a data set of 183,706 B-scans. A 3-channel ResNet-50 with Sigmoid activation adapted for multi-class multi-label detection (8 pathologies and 1 normal) was pre-trained on ImageNet images and was transfer trained with B-scans (224 x 224). In model 1, each B-scan is replicated to form 3-channel image. In model 2, each B-scan is combined with its two adjacent scans (above and below the corresponding B-scan) to create a localized-3D 3-channel image. In the 5-fold validation, if the probability of normal class was greater than 0.5, the B-scan was labeled as normal. If not, the B-scan was labeled as one or more of the pathologies.

Results : Figure 1 shows the average accuracy scores from 5-fold validation for both the models and the accuracies of all the labels increased in model 2 with the addition of localized 3D information. Figure 2 shows some of the examples where the proposed localized-3D information in model 2 improved the accuracy of the deduction.

Conclusions : The proposed model using localized-3D information performs better than the model using single B-scan. In this study, we developed a deep learning algorithm for multiple pathology detection that has the potential to function as a clinical decision support tool to improve the workflow efficiency.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure1. Average accuracy scores from 5-fold validation

 

Figure 2. (a) False positives and (b) false negatives in model 1 which were classified correctly in model 2; (c) false positives and (d) false negatives in both model 1 and model 2