Artificial intelligence deep learning algorithm for discriminating ungradable optical coherence tomography...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Spectral-Domain Optical Coherence Tomography (SDOCT) provides crucial 3-D imaging of the optic nerve head (ONH).
• Current quality assessment relies on signal strength (SS), which is insufficient for detecting various artifacts.
• Artifacts like motion, blurriness, and off-centration compromise SDOCT scan reliability for disease detection.

Purpose of the Study:

• To develop and validate a deep learning system (DLS) for automated filtering of ungradable SDOCT ONH volumes.
• To improve the accuracy and efficiency of SDOCT image quality control.
• To establish an automated method for identifying SDOCT scans with artifacts that affect ONH structure assessment.

Main Methods:

• A 3-D deep learning system (DLS) utilizing squeeze-and-excitation ResNeXt blocks was developed.
• 5599 SDOCT ONH volumes were used for training and primary validation.
• External validation was performed on two independent datasets (711 and 298 volumes).
• Ungradable scans were defined by low signal strength or artifacts impacting critical measurement areas.

Main Results:

• The DLS demonstrated good performance in both primary and external validation datasets.
• Key performance metrics included area under the receiver operating characteristic curve (AUC), sensitivity, specificity, and accuracy.
• Gradient-weighted class activation maps were used to visualize DLS decision-making.

Conclusions:

• The developed DLS effectively filters ungradable SDOCT volumes, addressing limitations of traditional signal strength assessment.
• Automated quality control using this DLS can enhance the efficiency and accuracy of SDOCT volumetric scan analysis.
• This system has the potential to streamline the detection of optic nerve head diseases by ensuring reliable image data.