arXiv:2510.16146v2 Announce Type: replace 
Abstract: The limited availability of annotated data in medical imaging makes semi-supervised learning increasingly appealing for its ability to learn from imperfect supervision. Recently, teacher-student frameworks have gained popularity for their training benefits and robust performance. However, jointly optimizing the entire network can hinder convergence and stability, especially in challenging scenarios. To address this for medical image segmentation, we propose DuetMatch, a novel dual-branch semi-supervised framework with asynchronous optimization, where each branch optimizes either the encoder or decoder while keeping the other frozen. To improve consistency under noisy conditions, we introduce Decoupled Dropout Perturbation, enforcing regularization across branches. We also design Pair-wise CutMix Cross-Guidance to enhance model diversity by exchanging pseudo-labels through augmented input pairs. To mitigate confirmation bias from noisy pseudo-labels, we propose Consistency Matching, refining labels using stable predictions from frozen teacher models. Extensive experiments on benchmark brain MRI segmentation datasets, including ISLES2022 and BraTS, show that DuetMatch consistently outperforms state-of-the-art methods, demonstrating its effectiveness and robustness across diverse semi-supervised segmentation scenarios.

DuetMatch هو إطار جديد ثنائي الفرع شبه مشرف مصمم لتجزئة صور الرنين المغناطيسي للدماغ، حيث يتناول التحديات التي تطرحها محدودية توفر البيانات المعلّمة. يستخدم تحسينًا غير متزامن، مما يسمح لفرع واحد بتحسين المشفر بينما يبقى الآخر ثابتًا. يقدم الإطار اضطراب الإسقاط المنفصل لتحسين الاتساق في الظروف المزعجة وCross-Guidance CutMix بالزوج لتعزيز تنوع النموذج من خلال تبادل التسميات الزائفة.

DuetMatch es un nuevo marco semi-supervisado de doble rama diseñado para la segmentación de resonancias magnéticas cerebrales, abordando los desafíos que plantea la disponibilidad limitada de datos anotados. Emplea optimización asíncrona, permitiendo que una rama optimice el codificador mientras la otra permanece congelada. El marco introduce Perturbación de Dropout Desacoplada para mejorar la consistencia en condiciones ruidosas y Cross-Guidance CutMix por pares para aumentar la diversidad del modelo mediante el intercambio de pseudoetiquetas.

DuetMatch est un nouveau cadre semi-supervisé à double branche conçu pour la segmentation des IRM cérébrales, répondant aux défis posés par la disponibilité limitée de données annotées. Il utilise une optimisation asynchrone, permettant à une branche d'optimiser l'encodeur tandis que l'autre reste figée. Le cadre introduit la Perturbation de Dropout Découplée pour améliorer la cohérence dans des conditions bruyantes et le Cross-Guidance CutMix par paires pour renforcer la diversité du modèle grâce à l'échange de pseudo-étiquettes.

DuetMatch is a novel dual-branch semi-supervised framework designed for brain MRI segmentation, addressing the challenges posed by limited annotated data. It employs asynchronous optimization, allowing one branch to optimize the encoder while the other remains frozen. The framework introduces Decoupled Dropout Perturbation for improved consistency under noisy conditions and Pair-wise CutMix Cross-Guidance to enhance model diversity through the exchange of pseudo-labels.

DuetMatch: Harmonizing Semi-Supervised Brain MRI Segmentation via Decoupled Branch Optimization

arXiv:2506.03942v3 Announce Type: replace 
Abstract: Deep neural networks for medical image segmentation are often overconfident, compromising both reliability and clinical utility. In this work, we propose differentiable formulations of marginal L1 Average Calibration Error (mL1-ACE) as an auxiliary loss that can be computed on a per-image basis. We compare both hard- and soft-binning approaches to directly improve pixel-wise calibration. Our experiments on four datasets (ACDC, AMOS, KiTS, BraTS) demonstrate that incorporating mL1-ACE significantly reduces calibration errors, particularly Average Calibration Error (ACE) and Maximum Calibration Error (MCE), while largely maintaining high Dice Similarity Coefficients (DSCs). We find that the soft-binned variant yields the greatest improvements in calibration, over the Dice plus cross-entropy loss baseline, but often compromises segmentation performance, with hard-binned mL1-ACE maintaining segmentation performance, albeit with weaker calibration improvement. To gain further insight into calibration performance and its variability across an imaging dataset, we introduce dataset reliability histograms, an aggregation of per-image reliability diagrams. The resulting analysis highlights improved alignment between predicted confidences and true accuracies. Overall, our approach not only enhances the trustworthiness of segmentation predictions but also shows potential for safer integration of deep learning methods into clinical workflows. We share our code here: https://github.com/cai4cai/Average-Calibration-Losses

قدمت الأبحاث الأخيرة صيغة تفاضلية لخطأ المعايرة الهامشي L1 (mL1-ACE) كخسارة مساعدة لشبكات الأعصاب العميقة في تقسيم الصور الطبية، مما يعالج مشكلة الثقة الزائدة في التنبؤات. أظهرت الدراسة أن دمج mL1-ACE يقلل بشكل كبير من أخطاء المعايرة عبر أربعة مجموعات بيانات، بما في ذلك ACDC وBraTS، مع الحفاظ على معاملات تشابه Dice العالية.

Investigaciones recientes han presentado una formulación diferenciable del error de calibración marginal L1 (mL1-ACE) como una pérdida auxiliar para redes neuronales profundas en la segmentación de imágenes médicas, abordando el problema de la sobreconfianza en las predicciones. El estudio demostró que la incorporación de mL1-ACE reduce significativamente los errores de calibración en cuatro conjuntos de datos, incluidos ACDC y BraTS, manteniendo altos coeficientes de similitud de Dice.

Des recherches récentes ont introduit une formulation différentiable de l'erreur de calibration marginale L1 (mL1-ACE) comme perte auxiliaire pour les réseaux de neurones profonds dans la segmentation d'images médicales, abordant le problème de la surconfiance dans les prédictions. L'étude a démontré que l'incorporation de mL1-ACE réduit considérablement les erreurs de calibration sur quatre ensembles de données, dont ACDC et BraTS, tout en maintenant des coefficients de similarité de Dice élevés.

Recent research has introduced a differentiable formulation of marginal L1 Average Calibration Error (mL1-ACE) as an auxiliary loss for deep neural networks in medical image segmentation, addressing the issue of overconfidence in predictions. The study demonstrated that incorporating mL1-ACE significantly reduces calibration errors across four datasets, including ACDC and BraTS, while maintaining high Dice Similarity Coefficients.

DuetMatch: Harmonizing Semi-Supervised Brain MRI Segmentation via Decoupled Branch Optimization

Was this article worth reading? Share it

LucidQuery AI

Augmeta

Attentive AI