Multi-Resolution Feature Stem for Diabetic Retinopathy lesion segmentation

2026-07-09

AI Summary

This study addresses the challenge posed by the significant scale variation of lesions in diabetic retinopathy (DR), where a single input resolution struggles to simultaneously capture both minute and large pathological structures. The authors systematically demonstrate that while high-resolution inputs benefit the detection of microaneurysms, they may degrade segmentation performance for extensive hemorrhagic regions. To reconcile this trade-off, they propose a parameter-efficient multi-resolution input-level pyramid architecture integrated into the UNet++ backbone, which enables parallel fusion of multi-scale features to balance fine-grained detail preservation with contextual awareness. Extensive experiments across multiple mainstream architectures and input resolutions (512×512 and 1024×1024) show that the proposed method significantly enhances overall DR lesion segmentation accuracy while maintaining model compactness, thereby validating the efficacy and necessity of a multi-resolution strategy.

Abstract

Diabetic Retinopathy (DR) is a leading cause of preventable blindness worldwide, requiring automated lesion segmentation using deep learning models for early detection and monitoring. However, DR lesions vary dramatically in size from tiny microaneurysms to large hemorrhages and exudates. This variability creates conflicting demands on the model architecture and input resolution, posing a challenge for effective design. This work investigates the impact of input resolution on different lesion types. Through systematic experimentation with multiple architectures (U-Net, UNet++, Vision Transformers, DeepLabV3+) at $512 \times 512$ and $1024 \times 1024$ resolutions, we identify a critical, counter-intuitive phenomenon where increasing input resolution has opposing effects on different lesion types. We demonstrate that while higher resolution is essential for resolving fine-grained microaneurysms, it can unexpectedly degrade performance on larger hemorrhages. This finding challenges the common assumption that higher resolution is uniformly beneficial. To address this, we propose a novel Multi-Resolution Feature Stem, an input-level pyramid integrated with a UNet++ backbone. This architecture processes multiple scales in parallel, capturing fine-grained details without sacrificing contextual information. This work contributes crucial empirical evidence of this complex, resolution-dependent behavior and a practical, parameter-efficient architecture that successfully resolves this trade-off.

Problem

• Diabetic Retinopathy
• lesion segmentation
• input resolution
• multi-scale lesions
• resolution trade-off

Innovation

• Multi-Resolution Feature Stem
• Diabetic Retinopathy
• Lesion Segmentation
• Input Resolution
• UNet++