Source-linked AI summary
Bayesian-Optimized Superpixel-GrabCut for Traceable Optic Disc Segmentation
Shraddha Changune, Vivek Noel Soren, Gautam Das, Tapan Kumar Gandhi
TL;DR
Optic disc segmentation needs accuracy without sacrificing inference-stage traceability, a limitation of opaque deep learning approaches. The paper combines explicit superpixel, morphological, GrabCut, and elliptical-shape operations with Bayesian hyperparameter optimization, achieving a Dice coefficient of 0.9536 on Drishti-GS while retaining deterministic algorithmic transparency.
Problem
Deep learning approaches for optic disc segmentation lack inference-stage mathematical traceability needed for clinical auditing and failure analysis.
Method
The paper uses a traceable pipeline combining superpixel seed selection and scoring, morphological refinement, GrabCut delineation, elliptical regularization, and Bayesian hyperparameter optimization.
Results
0.9536 Dice coefficient on Drishti-GS, matching the best traditional approaches while achieving competitive performance against deep learning baselines.
Takeaways & Limitations
The framework provides a deterministic, mathematically traceable alternative to black-box architectures for regulatory review and debugging.
Abstract
from arXiv · showhide
Optic disc (OD) segmentation is essential for diagnosing ophthalmic pathologies from retinal fundus images. However, prevailing deep learning approaches operate as opaque black boxes, lacking the inference-stage mathematical traceability--a critical requirement for algorithmic auditing and failure analysis in clinical workflows. This paper presents a fully algorithmically traceable and trainable segmentation pipeline that jointly combines superpixel decomposition, hybrid brightness-proximity superpixel scoring, morphological regularization, iterative GrabCut refinement, and elliptical shape fitting. The hyperparameter optimization is formulated as an objective function and solved via Bayesian optimization to eliminate manual parameter tuning. A quantitative evaluation on the Drishti-GS dataset demonstrates that our method achieves a Dice coefficient of 0.9536, matching state-of-the-art performance. By maintaining explicit mathematical transparency across all processing stages, our framework offers a deterministic, traceable alternative to black-box architectures for medical review and debugging.
I. Introduction
The paper proposes a trainable, mathematically traceable optic disc segmentation pipeline as an auditable alternative to opaque deep learning models. It combines explicit segmentation components with Bayesian hyperparameter optimization to retain transparent inference while targeting competitive performance.
- Motivation: Deep learning models can hinder clinical trust because their decision pathways are difficult to interpret, whereas traditional methods expose mathematical constraints.The paper contrasts opaque learned parameter spaces with explicit objectives and energy minimization.
- Proposed approach: The proposed pipeline segments optic discs without deep learning networks while remaining fully trainable and traceable.It is designed to delineate the optic disc from fundus images using an explicitly specified processing pipeline.
- Proposed approach: The framework integrates superpixel seed selection and scoring, GrabCut boundary refinement, and elliptical shape regularization.These components provide an auditable alternative to opaque black-box models.
- Optimization: Hyperparameter tuning is formulated as a global objective and solved with Bayesian optimization instead of manual parameter selection.The optimization is intended to select parameters systematically within the traceable pipeline.
A. Segmentation Pipeline
The segmentation pipeline enhances the red channel, decomposes the image into SLIC superpixels, and selects candidate regions using complementary brightness and proximity cues. Morphological closing and hole filling then regularize the initial mask before later refinement stages.
- Preprocessing: The red channel is selected because it typically provides strong contrast between the pale optic disc and surrounding tissue.CLAHE is applied to improve separability under local intensity variation.
- Superpixel Decomposition: CLAHE-enhanced images are converted to three channels and partitioned into superpixels with SLIC.SLIC groups perceptually similar pixels while preserving boundaries relevant to the optic disc margin.
- Superpixel Scoring and Selection: Superpixel selection combines brightness and spatial proximity to the brightest superpixel as complementary cues.The brightest superpixel provides an anchor, while proximity favors spatially coherent candidate regions.
- Superpixel Scoring and Selection: The top k superpixels with the smallest combined scores are selected, with the brightest seed forcibly included when necessary.The selected superpixels are united to form the initial optic disc mask.
- Morphological Refinement: Morphological closing with an elliptical structuring element fills narrow vessel-induced gaps and smooths the boundary.Small holes below the tunable threshold τhole are subsequently removed by morphological hole filling.
Morphological Refinement:
Morphological refinement cleans the initial OD mask and selects the component associated with the seed superpixel before boundary refinement.
- Morphological Refinement:: Morphological hole filling removes small holes below the tunable threshold τhole after closing smooths the mask.The threshold ranges from 50 to 500 pixels.
- Morphological Refinement:: Connected components are analyzed to retain the component containing the seed superpixel.
- Morphological Refinement:: If the seed lies outside all components, the largest component is selected as the coarse localization mask.The resulting mask is passed to GrabCut for precise boundary delineation.
- Morphological Refinement:: GrabCut receives the localization mask and uses a tunable band around it to define probable background and foreground regions.
GrabCut Boundary Refinement:
GrabCut refines the coarse localization mask by constructing a trimap and iteratively optimizing an energy function for boundary delineation.
- GrabCut Boundary Refinement:: Pixels outside the localization mask and farther than β from the foreground are assigned to probable background.
- GrabCut Boundary Refinement:: Pixels outside the localization mask within distance β of the foreground form the unknown region.
- GrabCut Boundary Refinement:: The resulting trimap guides GrabCut, which combines Gaussian-mixture color likelihoods and contrast-sensitive smoothness through iterative min-cut/max-flow optimization.The refinement runs for Tgc iterations and produces Mgc.
- GrabCut Boundary Refinement:: A least-squares ellipse is fitted to the refined mask contour using its center, axis lengths, and orientation.
Ellipse Regularization:
The pipeline combines guided superpixel selection, morphological cleanup, connected-component localization, GrabCut refinement, and ellipse-based output construction, with Bayesian optimization selecting parameters.
- B. Hyperparameter Optimization: Bayesian optimization minimizes negative mean Dice over a defined parameter space using a Gaussian-process surrogate and Expected Improvement.The parameter configuration with the lowest observed objective is returned after the predefined iteration budget.
- Ellipse Regularization:: The algorithm extracts the red channel, applies CLAHE, and decomposes the enhanced image into SLIC superpixels.
- Ellipse Regularization:: Superpixels are scored by normalized distance and brightness, and the top k lowest-scoring regions are combined into the initial mask.The seed superpixel is forcibly included when necessary.
- Ellipse Regularization:: Morphological closing, hole removal, and seed-based connected-component selection produce the localization mask before GrabCut.
- Ellipse Regularization:: Least-squares ellipse fitting rasterizes the refined contour, with a convex-hull fallback when fitting fails.
- B. Hyperparameter Optimization: Optimization is performed during offline model selection, while final inference uses fixed parameters and remains deterministic.
A. Dataset
Evaluation uses the publicly available Drishti-GS retinal fundus dataset with expert pixel-level annotations and the official training-testing split.
- A. Dataset: Drishti-GS contains 101 color retinal fundus photographs at 2896 × 1944-pixel resolution.
- A. Dataset: Expert annotations provide pixel-level masks for the background, optic disc, and optic cup.
- A. Dataset: The official split assigns 50 images to training and 51 images to testing.
- A. Dataset: The training set is further divided into 35 development-training images and 15 validation images for hyperparameter optimization.
B. Implementation details
The pipeline is implemented as a modular framework with 10 tunable parameters, optimized using Bayesian optimization rather than manual tuning.
- Implementation details: The modular framework contains a core segmentation routine and a hyperparameter optimizer operating on native-resolution images.Morphological operations use elliptical structuring elements.
- Implementation details: Ten tunable parameters are sampled from discrete integer sets or continuous real-valued intervals.
- Implementation details: Bayesian optimization uses a Gaussian-process surrogate with a Matérn 5/2 kernel and Expected Improvement acquisition.
- Implementation details: The optimization runs for 10 iterations, beginning with 2 random evaluations and continuing with 8 GP-guided evaluations.
C. Evaluation Metrics
Evaluation combines pixel-wise segmentation metrics with qualitative examples and component-wise analysis. The full pipeline reaches a Dice coefficient of 0.9536, while ablations expose the effects of individual modules.
- Evaluation metrics: The evaluation uses Dice, sensitivity, specificity, and precision derived from pixel-wise confusion-matrix counts.
- Quantitative evaluation: 0.9536 Dice is achieved by the full pipeline in the component-wise evaluation.
- Component-wise analysis: Removing GrabCut produces a Dice of 0.8742 and precision of 0.8922, the largest degradation among refinement-stage ablations.
- Qualitative evaluation: Figure 2 reports Dice scores of 0.972 and 0.897 for two qualitative Drishti-GS examples, with outputs progressing from SLIC decomposition to coarse and refined masks.
- Component-wise analysis: Disabling Bayesian optimization reduces Dice to 0.7105 and sensitivity to 0.5543.
- Component-wise analysis: Ellipse regularization lowers sensitivity to 0.8962 while increasing precision to 0.9869.
Comparison with State-of-the-Art Methods:
On Drishti-GS, the traceable pipeline achieves a Dice of 0.9536 while matching the best traditional approaches and supporting regulatory review through algorithmic traceability.
- Comparison with State-of-the-Art Methods: The method achieves a Dice of 0.9536 on the Drishti-GS dataset, competitive with the best traditional approaches.
- Comparison with State-of-the-Art Methods: The pipeline integrates SLIC superpixel selection, morphological refinement, GrabCut delineation, elliptical regularization, and Bayesian hyperparameter optimization.
- Comparison with State-of-the-Art Methods: The framework is presented as fully traceable for regulatory review, with component-wise ablation validating each module’s contribution.