Rethinking Convergence in Deep Learning: The Predictive-Corrective Paradigm for Anatomy-Informed Brain MRI Segmentation

Feifei Zhang, Zhenhong Jia, Sensen Song, Fei Shi, Dayong Ren

Keywords: Deep Learning, Medical Image Segmentation, Mamba

Abstract:

In medical image segmentation, although end-to-end deep learning has achieved substantial progress, obtaining accurate results typically requires many training iterations and large-scale annotated datasets, which limits efficiency and practicality in data-scarce clinical scenarios. To address this issue, we propose a Predictive–Corrective (PC) paradigm that decouples segmentation into a fast, anatomy-informed prediction stage followed by a focused refinement stage. Based on this paradigm, we develop PCMambaNet, which comprises two cooperative modules: a Predictive Prior Module (PPM) that generates a coarse anatomical approximation at low computational cost and injects symmetry priors via inter-hemispheric similarity and thresholding to highlight diagnostically relevant asymmetric regions, and a Corrective Residual Network (CRN) that models the residual error, concentrating capacity on refining challenging regions and delineating pathological boundaries.
Experiments on multiple high-resolution brain MRI benchmarks show that PCMambaNet attains competitive accuracy with relatively few training epochs and exhibits clear advantages in data-limited settings. Extended experiments further indicate that the proposed PC paradigm remains applicable to organs without strong left–right symmetry. Overall, this work demonstrates that explicitly incorporating anatomy-informed priors and decoupling prediction from refinement is an effective way to improve both training efficiency and data efficiency in medical image segmentation.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3434