ResUNet++ is built upon the Deep Residual U-Net (ResUNet) and UNet. However, ResUNet++ goes further, embedding:

Block diagram of ResUNet++
  • Residual Blocks: Ensuring information propagation across layers.
  • Squeeze and Excitation Blocks: Recalibrating feature responses for better representation.
  • ASPP (Atrous Spatial Pyramidal Pooling): Enlarging filter’s field-of-view to capture broader contexts.
  • Attention Blocks: Enhancing the relevance of feature maps in the decoding path.

Moreover, the architecture is uniquely structured with a stem block, three encoder blocks, three decoder blocks, and an ASPP. The output from the decoder undergoes another ASPP, followed by a 1×1 convolution, culminating in a segmentation map.

Digging Deeper

  • Residual Units: With the challenges of training deeper neural networks, residual units make it easier. They carry information without degradation and minimize the parameters, thereby maintaining or boosting performance.
  • Squeeze and Excitation Units: This unit recalibrates feature responses. By squeezing (through global average pooling) and then exciting (capturing channel-wise dependencies), it ensures the network focuses on the most relevant features.
  • Atrous Spatial Pyramidal Pooling (ASPP): Inspired by spatial pyramidal pooling, ASPP captures contextual information at varying scales. This is done using atrous convolutions at different rates, enabling precise multi-scale information capture.
  • Attention Units: Originally popular in Natural Language Processing, attention mechanisms dictate which parts of the input network should be prioritized. They are a great way to boost results and optimize computational resources.

Results on ResUNet++

ResUNet++ was rigorously tested against two significant datasets: Kvasir-SEG and CVC-ClinicDB. These datasets were geared towards polyp image segmentation. ResUNet++ results were compared with those obtained from U-Net and ResUNet.

Key Observations:

ResUNet++ consistently outperformed its counterparts, particularly in metrics crucial for segmentation tasks such as dice coefficient and mIoU. The qualitative results of segmentation masks revealed that ResUNet++ produced more accurate and refined masks than others.

Qualitative results on Kvasir-SEG dataset

Discussion and Conclusion

ResUNet++ has proven its efficacy in the realm of medical image segmentation. Its architectural nuances, like the attention mechanism and the squeeze and excitation units, give it an edge over its predecessors. It does not just achieve better segmentation but does so with optimized computational cost.

For anyone diving into medical image segmentation, ResUNet++ undoubtedly offers a robust framework. Its superior performance holds promise for real-world clinical applications, ensuring precise and effective results.


  1. Jha, D., Smedsrud, P. H., Johansen, D., de Lange, T., Johansen, H. D., Halvorsen, P., & Riegler, M. A. (2021). A comprehensive study on colorectal polyp segmentation with ResUNet++, conditional random field and test-time augmentation. IEEE journal of biomedical and health informatics25(6), 2029-2040.
  2. Jha, D., Smedsrud, P. H., Riegler, M. A., Johansen, D., De Lange, T., Halvorsen, P., & Johansen, H. D. (2019, December). Resunet++: An advanced architecture for medical image segmentation. In 2019 IEEE international symposium on multimedia (ISM) (pp. 225-2255). IEEE.

Leave a Reply

Your email address will not be published. Required fields are marked *