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

## TL;DR

This study uses deep learning and explainability analysis to automatically segment macular holes and cysts in OCT images, ensuring accurate diagnosis and treatment planning.

## Contribution

The novel contribution is the use of GradientSHAP-based metrics to quantitatively evaluate model explainability in OCT segmentation.

## Key findings

- The model achieved high Dice/IoU scores (0.94/0.91 for holes and 0.87/0.81 for cysts) with good boundary accuracy and calibration.
- Three distinct SHAP-based explainability regimes were identified, highlighting different focus patterns in lesion segmentation.
- In prediction mode, the model showed improved agreement with segmentations, particularly for macular holes.

## Abstract

Background: Optical coherence tomography (OCT) can quantify the morphology and dimensions of a macular hole for diagnosis and treatment planning. Objective: The aim of this study was to perform automatic segmentation of macular holes (MHs) and cysts from OCT macular volumes using a deep learning-based model and to quantitatively evaluate decision reliability using the model’s focus regions and GradientSHAP-based explainability. Methods: In this study, we automatically segmented MHs and cysts in OCT images from the open-access OIMHS dataset. The dataset comprises 125 eyes from 119 patients and 3859 OCT B-scans. OCT B-scan slices were input to a UNet-48-based model with a 2.5D stacking strategy. Performance was evaluated using Dice and intersection-over-union (IoU), boundary accuracy was evaluated using the 95th-percentile Hausdorff distance (HD95), and calibration was evaluated using the expected calibration error (ECE). Explainability was quantified from GradientSHAP maps using lesion coverage and spatial focus metrics: Attribution Precision in Lesion (APILτ), which is the proportion of attributions (SHAP contributions) falling inside the lesion; Attribution Recall in Lesion (ARILτ), which is the proportion of the true lesion covered by the attributions; and leakage (Leakτ = 1 − APILτ), which is the proportion of attributions falling outside the lesion. Spatial focus was monitored using the center-of-mass distance (COM-dist), which is the Euclidean distance between the attribution center and the segmentation center. All metrics were calculated using the top τ% of the pixels with the highest SHAP values. SHAP features were clustered using PCA and k-means. Explanations were calculated using the clinical mask in ground truth (GT) mode and the model segmentation in prediction (Pred) mode. Results: The Dice/IoU values for holes and cysts were 0.94/0.91 and 0.87/0.81, respectively. Across lesion classes, HD95 = 6 px and ECE = 0.008, indicating good boundary accuracy and calibration. In GT mode (τ = 20), three regimes were observed: (i) retina-dominant: high ARIL (hole: 0.659; cyst: 0.654), high Leak (hole: 0.983; cyst: 0.988), and low COM-dist (hole: 7.84 px; cyst: 6.91 px), with the focus lying within the retina and largely confined to the retinal tissue; (ii) peri-lesional: highest ARIL (hole: 0.684; cyst: 0.719), relatively lower Leak (hole: 0.917; cyst: 0.940), and medium/high COM-dist (hole: 16.22 px; cyst: 10.17 px), with the focus located around the lesion; (iii) narrow-coverage: primarily seen for cysts in GT mode (ARIL: 0.494; Leak: 1.000; COM-dist: 52.02 px), with markedly reduced coverage. In Pred mode, the ARIL20 for holes increased in the retina-dominant cluster (0.758) and COM-dist decreased (6.24 px), indicating better agreement with the model segmentation. Conclusions: The model exhibited high accuracy and good calibration for MH and cyst segmentation in OCT images. Quantitative characterization of SHAP validated the model results. In the clinic, peri-lesion and narrow-coverage conditions are the key situations that require careful interpretation.