2MF-Net: 3D Cardiac Keypoint Detection with Multi-scale and Multi-dimension Feature Fusion Net

Cardiac keypoint detection is of significant importance in cardiac quantitative analysis, disease diagnosis, and treatment. To address the challenge of precise localization of cardiac keypoints in CT images, which are often small in proportion and closely resemble the density of surrounding structures, this study proposes a 3D cardiac keypoint detection network (2MF-Net) with multi-scale and multi-dimension feature fusion. The Multi-scale Feature Fusion Optimization module (MFFO) initially fuses feature maps of different scales from the decoder in parallel. After enhancing the features of each scale’s fusion results, a second fusion is performed to supplement information lost during the sampling process, thereby enhancing the network’s capability to express semantic information. The Multi-dimension Feature Enhancement Attention (MFEA) decomposes the feature maps into directional information across height, width, and depth dimensions. It then undertakes three rounds of fusion on these multi-dimensional features, refining their positional information and diminishing interference from structurally similar densities. Evaluation of 2MF-Net performance on a 3D cardiac keypoint dataset, using a two-stage experimental approach, yielded a Mean Radial Error (MRE) of 4.34mm and a Success Detection Rate (SDR) of 97.00%. These results underscore the efficacy of 2MF-Net in 3D cardiac keypoint detection tasks.