Objective:
Modeling of the ophthalmologist's decision-making process for activity detection in neovascular age-related macular degeneration with a multi-task convolutional deep neuronal network which takes intra- and subretinal fluid into account.
Design:
A cohort study to evaluate the multi-task deep learning model for activity detection.
Participants:
n = 70 patients (46 female, 24 male) attended the University Eye Hospital Tuebingen between 21.2.2018 and 27.6.2018. 3762 optical coherence tomography B-scans (right eye: 2011, left eye: 1751) were acquired from them with Heidelberg Spectralis, Heidelberg, Germany.
Methods:
B-scans were graded by a retina specialist and an ophthalmology resident, and then used to develop a multi-task deep learning model to concurrently predict disease activity in neovascular age-related macular degeneration along with the presence of sub- and intraretinal fluid.
Main outcome measures:
Performance metrics compared to single-task networks, visualization of the representation driving the DNN-based decisions using t-distributed stochastic neighbor embedding and analysis of the model's decisions via clinically validated saliency mapping techniques.
Results:
The multi-task model surpassed single-task networks in accuracy for activity detection. Visualizations via t-distributed stochastic neighbor embedding and saliency maps highlighted that the network's decisions for activity of neovascular age-related macular degeneration were based on the presence of sub- and intraretinal fluids, the optical coherence tomography characteristics used for treatment decision in clinical routine.
Conclusion:
Multi-task learning increases the performance of neuronal networks for predicting disease activity, while providing clinicians with an easily accessible decision control, which resembles human reasoning.