Source-linked AI summary
Bi-Directional ConvLSTM U-Net with Densley Connected Convolutions
Reza Azad, Maryam Asadi-Aghbolaghi, Mahmood Fathy, Sergio Escalera
TL;DR
Medical image segmentation requires pixel-level labels, while annotated datasets are limited. BCDU-Net extends U-Net with BConvLSTM skip connections, dense convolutions, and batch normalization. Evaluations on three datasets report state-of-the-art performance, and batch normalization reduced ISIC convergence from about 200 to about 30 epochs.
Problem
Medical image segmentation requires pixel-level labels, but large annotated datasets are unavailable.
Method
BCDU-Net extends U-Net by combining encoder and decoder features with BConvLSTM and reusing features through dense convolutions, with batch normalization in decoding.
Results
BCDU-Net outperforms state-of-the-art alternatives across retinal vessel, skin lesion, and lung segmentation evaluations.
Takeaways & Limitations
The proposed combination produces more precise segmentation results and substantially faster convergence.
Abstract
from arXiv · showhide
In recent years, deep learning-based networks have achieved state-of-the-art performance in medical image segmentation. Among the existing networks, U-Net has been successfully applied on medical image segmentation. In this paper, we propose an extension of U-Net, Bi-directional ConvLSTM U-Net with Densely connected convolutions (BCDU-Net), for medical image segmentation, in which we take full advantages of U-Net, bi-directional ConvLSTM (BConvLSTM) and the mechanism of dense convolutions. Instead of a simple concatenation in the skip connection of U-Net, we employ BConvLSTM to combine the feature maps extracted from the corresponding encoding path and the previous decoding up-convolutional layer in a non-linear way. To strengthen feature propagation and encourage feature reuse, we use densely connected convolutions in the last convolutional layer of the encoding path. Finally, we can accelerate the convergence speed of the proposed network by employing batch normalization (BN). The proposed model is evaluated on three datasets of: retinal blood vessel segmentation, skin lesion segmentation, and lung nodule segmentation, achieving state-of-the-art performance.
1. Introduction
Medical image segmentation is important for diagnosis and treatment, but it requires pixel-level labels and large annotated datasets are scarce. The paper proposes BCDU-Net, extending U-Net with bidirectional ConvLSTM skip connections and densely connected convolutions.
- Medical image segmentation supports diagnosis and treatment by providing pixel-level interpretations of medical images.
- Deep learning networks achieve strong medical-imaging results, but segmentation requires pixel-level labels rather than image-level labels.
- U-Net uses encoding and decoding paths to extract feature maps and produce segmentation maps at the input resolution.
- BCDU-Net replaces U-Net skip-connection concatenation with BConvLSTM to combine encoding and decoding features nonlinearly.Encoding features provide higher resolution, while decoding features contain more semantic information.
- Densely connected convolutions are used to improve feature propagation and reuse, while batch normalization accelerates convergence.The model is evaluated on retinal blood vessel, skin lesion, and lung nodule segmentation.
2. Related Work
Related work progresses from CNN-based segmentation to FCN, U-Net, and recurrent architectures that preserve spatial information or model label dependencies. The paper positions BCDU-Net as a U-Net extension with better reported segmentation performance and faster convergence from batch normalization.
- Other related approaches include CNNs for brain MRI and pancreas segmentation, recurrent networks for long-range label dependencies, and architectures such as ReSeg and DeepLab.
- CNN segmentation methods can lose spatial information when convolutional features enter fully connected layers.
- FCNs address this issue by replacing fully connected layers with convolutional and deconvolutional layers for end-to-end pixelwise segmentation.
- U-Net uses contracting and expanding paths, works with few training samples, and combines global location with contextual information.
- BCDU-Net extends U-Net and is reported to outperform state-of-the-art alternatives, while batch normalization significantly affects convergence speed.
3. Proposed Method
BCDU-Net combines U-Net’s encoder–decoder structure with dense convolutions in the final encoding layer, BConvLSTM skip connections, and batch normalization in decoding. These components promote feature reuse, combine local and semantic information, and stabilize training.
- BCDU-Net combines U-Net, bidirectional ConvLSTM, and densely connected convolutions.
- Dense Convolutions: Dense convolutional blocks in the final encoding layer receive concatenated outputs from all preceding blocks, enabling feature reuse and information flow.
- Decoding Path: The decoding path upsamples previous-layer features, while BConvLSTM combines them with corresponding encoder features instead of simple concatenation.Encoder features have higher resolution, whereas decoder features contain more semantic information.
- Batch Normalization: Batch normalization processes up-convolution outputs to standardize layer inputs and accelerate training.
- Bi-Directional ConvLSTM: ConvLSTM incorporates convolutions into input-to-state and state-to-state transitions to address spatial-correlation limitations of standard LSTM.
- Bi-Directional ConvLSTM: BConvLSTM uses forward and backward ConvLSTMs to model dependencies in both directions and combine their states nonlinearly.Its output represents bidirectional spatio-temporal information.
4. Experimental Results
BCDU-Net was evaluated on DRIVE, ISIC 2018, and lung segmentation datasets, with comparisons against state-of-the-art methods and analyses of convergence and architectural components. Across experiments, dense connections, BConvLSTM skip connections, and batch normalization were associated with improved segmentation performance or faster convergence.
- Cross-dataset evaluation: BCDU-Net outperformed state-of-the-art alternatives on DRIVE, ISIC 2018, and lung segmentation benchmarks.The evaluations used retinal blood vessel, skin lesion, and lung segmentation tasks.
- Dense connections: The d = 3 configuration with dense connections performed better than the configuration without dense blocks on DRIVE and ISIC.The lung experiments likewise report better performance for the network with dense connections.
- Convergence analysis: After the 30th epoch, BCDU-Net was almost converged on DRIVE, while ISIC convergence occurred after 40 epochs.The DRIVE validation accuracy initially exceeded training accuracy, which the authors relate partly to the dataset's small size and validation procedure.
- ISIC 2018 Dataset: 0.936 JS was achieved by BCDU-Net on ISIC, compared with 0.802 for the best ISIC challenge result.The paper reports this as a substantial gap over the challenge result.
- BConvLSTM ablation: BConvLSTM skip connections improved the standard U-Net results by combining encoded local-information features with decoded semantic-information features.The proposed combination replaces simple concatenation with a nonlinear feature-combination mechanism.
- Batch normalization: BN reduced ISIC convergence from 200 epochs to about 30 epochs, yielding 6.6 times faster convergence and improved BCDU-Net accuracy.The authors attribute this behavior to standardization of input distributions and a small regularization effect.
5. Conclusion
BCDU-Net combines BConvLSTM skip connections and densely connected convolutional blocks to capture more discriminative information for medical image segmentation. Across three public benchmark datasets, it produced more precise segmentation and achieved about sixfold speedup through batch normalization.
- BCDU-Net integrates BConvLSTM skip connections and densely connected convolutional blocks for medical image segmentation.These modifications capture more discriminative information and improve segmentation precision.
- About sixfold speedup was achieved by applying batch normalization after the up-convolutional layer.
- Experiments on three public benchmark datasets showed high semantic-segmentation gains relative to state-of-the-art alternatives.