BiO-Net: Learning Recurrent Bi-directional Connections for Encoder-Decoder Architecture (2007.00243v2)

Abstract: U-Net has become one of the state-of-the-art deep learning-based approaches for modern computer vision tasks such as semantic segmentation, super resolution, image denoising, and inpainting. Previous extensions of U-Net have focused mainly on the modification of its existing building blocks or the development of new functional modules for performance gains. As a result, these variants usually lead to an unneglectable increase in model complexity. To tackle this issue in such U-Net variants, in this paper, we present a novel Bi-directional O-shape network (BiO-Net) that reuses the building blocks in a recurrent manner without introducing any extra parameters. Our proposed bi-directional skip connections can be directly adopted into any encoder-decoder architecture to further enhance its capabilities in various task domains. We evaluated our method on various medical image analysis tasks and the results show that our BiO-Net significantly outperforms the vanilla U-Net as well as other state-of-the-art methods. Our code is available at https://github.com/tiangexiang/BiO-Net.

• The paper presents BiO-Net, a novel U-Net variant that uses recurrent bi-directional skip connections to refine features without increasing parameter count.
• The paper demonstrates enhanced performance in medical imaging tasks, achieving superior Dice Coefficient and IoU in nuclei and EM membrane segmentation.
• The paper reveals that the recursive architecture effectively balances low-level and high-level feature extraction, optimizing accuracy and efficiency for complex vision tasks.

BiO-Net: Learning Recurrent Bi-directional Connections for Encoder-Decoder Architecture

The research introduces BiO-Net: a novel approach enhancing the U-Net model, widely recognized for its applications in medical image analysis tasks such as semantic segmentation, super-resolution, image denoising, and inpainting. Unlike previous U-Net variants that increase model complexity by incorporating additional parameters or modules, BiO-Net employs bi-directional recurrent skip connections in U-Net's encoder-decoder structure, optimizing performance without adding extra parameters.

Methodological Insights

BiO-Net maintains the same core architecture as U-Net but differentiates itself through the integration of bi-directional skip connections between the encoder and decoder components. These connections enable mutual feature reuse, facilitating a recursive learning process that enhances the network's ability to preserve both low-level visual and high-level semantic details.

1. Forward Skip Connections: These connections maintain the original U-Net approach of transferring features from encoder to decoder, preserving semantic scales across various network levels.
2. Backward Skip Connections: This key innovation passes features back from decoders to encoders, enriching the encoding process with high-level semantic insights gathered during decoding. The recursive architecture allows these processes to iterate multiple times, representing an O-shaped inference trajectory.
3. Recurrent Inferences: The recursive mechanism within BiO-Net allows the network to iteratively refine features without increasing the parameter count, bypassing the typical trade-off between performance and complexity.

The network architecture incorporates standard convolution, batch normalization, and non-linearity (ReLU) within its blocks, optimized for efficient feature extraction across both encoding and decoding stages.

Experimental Results

The efficacy of BiO-Net is demonstrated across several tasks:

1. Nuclei Segmentation: Evaluated on datasets such as MoNuSeg and TNBC, BiO-Net offers superior performance in terms of Dice Coefficient and IoU compared to U-Net and its variants, showing significant gains in segmentation accuracy and generalization capability without additional computational demands.
2. EM Membrane Segmentation: When applied to neuronal boundary prediction, BiO-Net maintains strong performance metrics, evidenced by superior Rand F-scores compared to leading U-Net variants.
3. Super Resolution: BiO-Net successfully enhances histopathological imaging processes, outperforming notable benchmarks such as FSRCNN and SRResNet in PSNR metrics, validating its versatility beyond segmentation tasks.

Implications and Future Directions

The results underscore BiO-Net's potential to advance medical imaging applications through more efficient and accurate processing of complex visual datasets. Its approach of reusing existing feature pathways establishes new precedents in network design, reducing memory footprint while maintaining high performance. This framework could expand into broader AI applications, whereby recursive architectures and efficient parameter utilization are critical.

Future work might explore the integration of BiO-Net with different encoder-decoder systems beyond medical imaging, as well as its adaptation to other complex vision tasks. The demonstrated ability to refine features recursively—without increasing network size—holds promise for further exploration in areas requiring detailed image processing with constrained computational resources.