MedElevator Войти в исследования
← Поиск публикаций

DOI 10.21037/qims-2026-1-0189

Медицинская библиотека · Quantitative imaging in medicine and surgery · 2026

A multimodal complementary feature fusion network for myocardial pathological segmentation in multi-sequence cardiac magnetic resonance images.

Zhang Chang-Jiang, Hong Qi-Bin, Zhou Min, Tang Fu-Qin, Cai Hai-Peng, Feng Li-Li

СтатьяПеревод в очереди

Аннотация

BACKGROUND

Automated myocardial pathology segmentation in multi-sequence cardiac magnetic resonance (CMR) imaging plays an indispensable role in assessing myocardial viability and managing patients with myocardial infarction (MI). However, precise delineation remains exceptionally challenging due to low soft tissue contrast, highly irregular pathological geometries, residual cross-modal spatial misalignments, and scarcity of labeled datasets. This study aimed to develop a highly accurate, stable, and unified single-stage deep learning framework to simultaneously segment the left ventricular (LV) blood pool, right ventricular (RV) blood pool, left ventricular normal myocardium (LVM), myocardial edema, and myocardial scars from multi-sequence CMR images.

METHODS

We propose the Multimodal CMR Image Pathological Segmentation Network (MCIPS-Net), an end-to-end framework built upon a five-layer U-Net architecture. To address inter-modal intensity distribution heterogeneity among balanced steady-state free precession (BSSFP), late gadolinium enhancement (LGE), and T2-weighted (T2) sequences, a plug-and-play multimodal feature fusion (MFF) module was introduced in the encoding path to independently extract and cascade modality-specific semantic features. Furthermore, an efficient channel attention (ECA)-based fusion module was embedded before upsampling to strengthen region of interest (ROI) focus. Training stability and performance under extreme class imbalance were optimized using a multi-scale deep supervision (DS) strategy and an adaptive composite loss function merging Dice and Focal losses. Center-cropped images (256×256 pixels) were expanded 15-fold via geometric and elastic data augmentations, and post-processing was executed to eliminate minor isolated prediction artifacts.

RESULTS

Evaluated on the MyoPS 2020 Challenge test set (20 cases, 72 slices), MCIPS-Net achieved an average Dice score of 0.655±0.238 for scar segmentation and 0.714±0.086 for edema and scar joint segmentation, yielding a comprehensive average Dice score of 0.685. Ablation studies revealed that the MFF module contributed the most significant gain, increasing the baseline average Dice score by 4.5% (from 0.574 to 0.619). On the official leaderboard, our single-stage method outperformed the leading advanced single-stage method by 0.7% while demonstrating performance equivalent to the official second-place multi-stage pipeline. Crucially, the model exhibited the lowest standard deviation (SD) in Dice scores among all participating top algorithms, highlighting its superior stability.

CONCLUSIONS

The proposed MCIPS-Net establishes a simple, efficient, and robust single-stage solution that bypasses complex localization networks or cumbersome model integration. By effectively capitalizing on cross-modal complementary features, this stable pipeline offers a high-performance clinical tool for precise infarct size quantification and myocardial viability assessment.

Для цитирования: Zhang Chang-Jiang, Hong Qi-Bin, Zhou Min, Tang Fu-Qin, Cai Hai-Peng, Feng Li-Li. A multimodal complementary feature fusion network for myocardial pathological segmentation in multi-sequence cardiac magnetic resonance images.. Quantitative imaging in medicine and surgery. 2026. DOI: 10.21037/qims-2026-1-0189.
Подготовка и перевод публикации
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…

История · было → стало

Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.

Тип публикации указан по данным источника. Уровень доказательности автоматически не присваивается.