Beyond Performance Metrics: Automatic Deep Learning Retinal OCT Analysis Reproduces Clinical Trial Outcome

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Macular telangiectasia type 2 (MacTel2) is a degenerative retinal disease.
• Accurate measurement of ellipsoid zone (EZ) defects is crucial for assessing MacTel2 progression and treatment efficacy.
• Conventional methods for measuring EZ defects can be time-consuming and subjective.

Purpose of the Study:

• To validate a fully automatic, deep learning-based segmentation algorithm for measuring EZ defect areas in MacTel2.
• To assess the algorithm's efficacy beyond standard metrics by evaluating its performance on a clinical trial's primary outcome.
• To determine the clinical applicability of automated segmentation in MacTel2 trials.

Main Methods:

• A phase 2 clinical trial dataset (NCT01949324) involving 92 eyes from 62 MacTel2 participants was used.
• A fully automatic, deep learning algorithm segmented EZ defect areas from spectral domain OCT images at baseline and 24 months.
• The change in EZ defect area was calculated and compared between treatment groups, mirroring the clinical trial's primary outcome analysis.

Main Results:

• The deep learning algorithm measured a change in EZ defect area of 0.072±0.035 mm² (P=0.021).
• This result was comparable to semiautomatic measurements by expert readers (0.065±0.033 mm², P=0.025).
• The algorithm successfully reproduced the statistically significant primary outcome of the clinical trial.

Conclusions:

• The fully automatic deep learning algorithm demonstrated accuracy comparable to expert semiautomatic segmentation for EZ defect areas.
• The algorithm reliably reproduced the clinical trial's primary outcome, confirming its clinical utility.
• Validating automated segmentation against primary clinical trial endpoints offers a robust measure of its real-world applicability in MacTel2 research.