Neural Netw. 2025 Nov 23;196:108329. doi: 10.1016/j.neunet.2025.108329. Online ahead of print.
ABSTRACT
Magnetic Resonance Imaging (MRI) of the heart provides precise visualization of cardiac anatomical structures, making accurate automatic segmentation highly valuable for assisting in the diagnosis of cardiovascular diseases. Numerous deep learning-based segmentation methods have been proposed to date. While these approaches have yielded promising results, most rely heavily on large-scale datasets with meticulously annotated labels-a process that is both labor-intensive and time-consuming. To address this limitation, we introduce a novel weakly semi-supervised segmentation framework, termed PDFMSeg. It requires only a small amount of sparsely labeled data along with a large volume of unlabeled data for training, significantly reducing the dependency on finely annotated datasets. Specifically, we designed the Frequency-Domain Pseudo-Label Dynamic Mixed Supervision (FPLMS) and Partial Dice Loss (LpDice). The FPLMS enhances boundary constraint capability by dynamically blending high- and low-frequency components from pseudo-labels derived from diverse sources. Meanwhile, LpDice enriches the form of scribble supervision by integrating Dice loss with an ignore mask and the log-cosh function, further improving segmentation performance. Experimental results demonstrate that PDFMSeg outperforms several state-of-the-art weakly semi-supervised methods. At a 10 % annotation ratio, PDFMSeg achieved the best performance on the public cardiac MRI datasets ACDC and MSCMR, with the following average scores: DSC (83.57 % and 74.29 %), JC (71.81 % and 59.09 %), and HD95 (13.29 and 31.66). Moreover, the model requires only 1.82M parameters and 2.32G FLOPs. These results confirm the effectiveness of PDFMSeg and underscore its potential for clinical application. Code and models are available at: https://github.com/labiip/PDFMSeg.
PMID:41308264 | DOI:10.1016/j.neunet.2025.108329