[Paper Review] Explainable Continuous-Time Mask Refinement with Local Self-Similarity Priors for Medical Image Segmentation

Rajdeep Chatterjee, Sudip Chakrabarty+1 more|arXiv (Cornell University)|Feb 28, 2026

Cutaneous Melanoma Detection and Management0 citations

TL;DR

The paper introduces LSS-LTCNet, a boundary-aware foot ulcer segmentation model that fuses Local Structural Similarity priors with a continuous-time Liquid Time-Constant refinement loop, achieving state-of-the-art boundary precision (HD95 8.91 px, Dice 86.96%) with 25.70M parameters.

ABSTRACT

Accurate semantic segmentation of foot ulcers is essential for automated wound monitoring, yet boundary delineation remains challenging due to tissue heterogeneity and poor contrast with surrounding skin. To overcome the limitations of standard intensity-based networks, we present LSS-LTCNet:an ante-hoc explainable framework synergizing deterministic structural priors with continuous-time neural dynamics. Our architecture departs from traditional black-box models by employing a Local Self-Similarity (LSS) mechanism that extracts dense, illumination-invariant texture descriptors to explicitly disentangle necrotic tissue from background artifacts. To enforce topological precision, we introduce a Liquid Time-Constant (LTC) refinement module that treats boundary evolution as an ODEgoverned dynamic system, iteratively refining masks over continuous time-steps. Comprehensive evaluation on the MICCAI FUSeg dataset demonstrates that LSS-LTCNet achieves state-of-the-art boundary alignment, securing a peak Dice score of 86.96% and an exceptional 95th percentile Hausdorff Distance (HD95) of 8.91 pixels. Requiring merely 25.70M parameters, the model significantly outperforms heavier U-Net and transformer baselines in efficiency. By providing inherent visual audit trails alongside high-fidelity predictions, LSS-LTCNet offers a robust and transparent solution for computer-aided diagnosis in mobile healthcare (mHealth) settings.

Motivation & Objective

• Address the challenge of precise wound boundary delineation under variable illumination and tissue heterogeneity.
• Integrate deterministic Local Structural Similarity (LSS) priors into early encoder features to anchor tissue boundaries.
• Introduce a Bottleneck refinement loop based on Liquid Time-Constant (LTC) continuous-time dynamics for iterative mask refinement.
• Provide inherent ante-hoc explainability by exposing LSS-derived tissue priors as visual audit trails.

Proposed method

• Additive Local Structural Similarity (LSS) Fusion to inject local patch-correlation statistics into the encoder.
• LSS Extractor computes patch-wise cosine similarity across local neighborhoods and constructs an LSS map mapped to the encoder via a small convolution.
• Liquid Time-Constant (LTC) refinement loop solves a continuous-time ODE at the bottleneck to iteratively update a refinement token guiding the decoder.
• Boundary Alignment Loss (BAL) using MSE between Sobel gradients of the predicted map and the LSS-derived mean channel to enforce topology-consistent boundaries.
• Deep supervision with auxiliary heads to stabilize training and promote multi-scale learning.

Experimental results

Research questions

• RQ1 Can deterministic LSS priors improve segmentation boundaries in foot ulcers beyond traditional CNN/ViT baselines?
• RQ2 Does the BAL term synergize LSS and LTC components to achieve superior boundary precision and Dice scores?
• RQ3 Does the framework provide inherent explainability through accessible LSS maps that align with tissue morphology?

Key findings

• LSS-LTCNet achieves a Dice score of 86.96% and IoU of 79.54% on the MICCAI FUSeg validation set.
• HD95 is 8.91 pixels, representing a 30% improvement over the next best baseline.
• Model has 25.70M parameters and demonstrates higher efficiency than heavier U-Net and transformer baselines.
• Ablation shows adding LSS and LTC without BAL reduces Dice to 76.18%, highlighting BAL’s crucial role.
• Inherent explainability is provided by the deterministic LSS maps (Mean, Max, Boundary Std) guiding segmentation.