Automated Cardiac Diagnosis Based on Weakly Supervised Learning

Semantic segmentation of medical images is a significant and challenging task, yet traditional fully supervised methods are difficult to achieve satisfactory results due to the scarcity and difficulty of image data annotation. Thus, this paper proposes a semi-supervised semantic segmentation method based on cross-pseudo-supervision and attention modulation. This method uses an EfficientNet as a feature encoder to extract multi-level features for feature fusion; in the encoding stage, it adds an over-parameterized convolution layer with depthwise separable convolution operation, replacing the original convolution method; in the decoding stage, it uses Unet decoder to process the multi-level feature signals, and introduces activation modulation and recalibration mechanisms to enhance the secondary features and suppress the most sensitive and least sensitive feature information. This method trains the model using a cross-pseudo-supervision approach, and achieves a Dice index of 86.7% on the ACDC 1/10 annotated dataset, which is 6% higher than the CPS method; after using 1/3 annotated data, it achieves results similar to the traditional fully supervised semantic segmentation task. Additionally, this method has training friendliness, and when reaching 85% Dice, the training speed is more than ten times faster than traditional fully supervised learning.