Mid-Net: Rethinking efficient network architectures for small-sample vascular segmentation

Deep learning-based medical image segmentation methods have demonstrated significant clinical applications. However, training these methods on small-sample vascular datasets remains challenging due to the scarcity of labeled data and severe category imbalance. To address this, this paper proposes Mid-Net, which fully exploits the often-overlooked feature representation potential of the middle-layer network through cross-layer guidance to improve model learning efficiency in data-constrained environments. Mid-Net consists of three core components: the encoding path, the guidance path, and the calibration path. In the encoding path, a feature pyramid structure with large kernel convolutions is used to extract semantic information at different scales. The guidance path combines the sensitivity of the shallow-layer network to spatial details with the global perceptual abilities of the deep-layer network to provide more discriminative guidance to the middle-layer network in a feature-decoupled manner. The calibration path further calibrates the spatial location information of the middle-layer network through end-to-end supervised learning. Experiments conducted on the publicly available retinal vascular datasets DRIVE, STARE, and CHASE_DB1, as well as coronary angiography datasets DCA1 and CHUAC, demonstrate that Mid-Net achieves superior segmentation results with lower computational resource requirements compared to state-of-the-art methods.