Two-level semi-supervised collaborative medical image segmentation with bidirectional knowledge exchange

Traditional co-training methods fail to leverage ensemble learning effectively, resulting in resource waste. To address this, we propose a two-level co-training structure. The first-level models follow a classical co-training approach, while the second-level models utilize the ensemble results of the first-level models as pseudo-labels. This design enables second-level models to achieve better segmentation performance than individual first-level models. However, we find that the performance of second-level models is constrained by the learning capacity of first-level models. To mitigate this, we introduce a bidirectional knowledge exchange strategy inspired by pix2pixHD, where features of the second-level models are fed back into the first-level models. This bidirectional knowledge exchange, integrated within the two-level co-training structure, forms a positive feedback loop that enhances the performance of both levels, resulting in superior segmentation results. Extensive experiments on multiple benchmark datasets demonstrate that our approach exhibits strong competitiveness against state-of-the-art methods.