R3Net: Recursive Residual Refinement Network Architecture for Decoder-Free Medical Image Segmentation.

Abstract

Background: Medical image segmentation methods based on encoder-decoder architectures often achieve high accuracy but typically require substantial computational resources and contain redundant parameters.

Purpose: This study aims to develop an efficient decoder-free segmentation framework that maintains competitive performance by strengthening the encoding process.

Methods: We propose R3Net, an encoder-only segmentation architecture based on a recursive residual refinement (R3) mechanism. By recursively reusing encoder stages and progressively fusing multiscale features through residual pathways, R3Net reconstructs high-resolution features without requiring a dedicated decoder.

Results: Experiments on three medical imaging modalities-cardiac MRI (Automated Cardiac Diagnosis Challenge), abdominal CT (Synapse), and thyroid ultrasound (Thyroid Nodule Multimodal Learning)-demonstrate that R3Net achieves segmentation performance comparable to representative encoder-decoder models while reducing the number of model parameters and computational complexity.

Conclusion: R3Net provides an effective decoder-free alternative for medical image segmentation, suggesting that competitive dense prediction can be achieved through recursive refinement within the encoder.