$$\alpha $$-UNet++: A Data-Driven Neural Network Architecture for Medical Image Segmentation

UNet++, an encoder-decoder architecture constructed based on the famous UNet, has achieved state-of-the-art results on many medical image segmentation tasks. Despite improved performance, UNet++ introduces densely connected decoding blocks, some of which, however, are redundant for a specific task. In this paper, we propose \(\alpha \)-UNet++ that allows us to automatically identify and discard redundant decoding blocks without the loss of precision. To this end, we design an auxiliary indicator function layer to compress the network architecture via removing a decoding block, in which all individual responses are less than a given threshold \(\alpha \). We evaluated the segmentation architecture obtained respectively for liver segmentation and nuclei segmentation, denoted by UNet++\(^C\), against UNet and UNet++. Comparing to UNet++, our UNet++\(^C\) reduces the parameters by 18.89% in liver segmentation and 34.17% in nuclei segmentation, yielding an average improvement of IoU by 0.27% and 0.11% on two tasks. Our results suggest that the UNet++\(^C\) produced by the proposed \(\alpha \)-UNet++ not only improves the segmentation accuracy slightly but also reduces the model complexity considerably.