Quantifying Explainability in OCT Segmentation of Macular Holes and Cysts: A SHAP-Based Coverage and Factor...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Optical coherence tomography (OCT) is crucial for quantifying macular hole (MH) morphology and dimensions, aiding diagnosis and treatment planning.
• Accurate segmentation of MHs and associated cysts in OCT volumes is essential for effective clinical management.

Purpose of the Study:

• To develop and evaluate a deep learning model for automatic segmentation of macular holes (MHs) and cysts in OCT macular volumes.
• To quantitatively assess the model's decision reliability using focus regions and GradientSHAP-based explainability.

Main Methods:

• A UNet-48-based deep learning model with a 2.5D stacking strategy was employed for segmenting MHs and cysts in the OIMHS OCT dataset.
• Performance was assessed using Dice, intersection-over-union (IoU), Hausdorff distance (HD95), and expected calibration error (ECE).
• GradientSHAP was utilized for explainability, quantifying lesion coverage (APILτ, ARILτ) and spatial focus (COM-dist).

Main Results:

• The model achieved high segmentation performance with Dice/IoU values of 0.94/0.91 for MHs and 0.87/0.81 for cysts.
• Excellent boundary accuracy (HD95 = 6 px) and calibration (ECE = 0.008) were demonstrated across lesion classes.
• GradientSHAP analysis revealed distinct explanation regimes (retina-dominant, peri-lesional, narrow-coverage), with improved agreement in prediction mode.

Conclusions:

• The deep learning model demonstrates high accuracy and good calibration for segmenting MHs and cysts in OCT images.
• Quantitative explainability using GradientSHAP validates the model's segmentation results and provides insights into its decision-making process.
• Peri-lesion and narrow-coverage scenarios require careful clinical interpretation, highlighting the utility of explainability in identifying challenging cases.