Source-linked AI summary
Deeply-Supervised CNN for Prostate Segmentation
Qikui Zhu, Bo Du, Baris Turkbey, Peter L . Choyke, Pingkun Yan
TL;DR
Prostate MRI segmentation is difficult because boundaries are unclear at the apex and base and prostate appearance varies across patients. The paper introduces a deeply supervised CNN that forwards early features to later stages while supervising hidden layers. Its average DSC reaches 0.885, above U-Net’s 0.865 and FCN’s 0.759.
Problem
Prostate MRI segmentation must handle unclear apex and base boundaries and substantial variation in prostate shape and texture across patients.
Method
The paper proposes a deeply supervised CNN that forwards early-stage features to later stages and adds supervised layers during end-to-end training.
Results
Average DSC values were 0.885 for the proposed model, 0.865 for U-Net, and 0.759 for FCN.
Takeaways & Limitations
The results support deeply supervised training as a viable approach for improving medical-image segmentation performance.
Abstract
from arXiv · showhide
Prostate segmentation from Magnetic Resonance (MR) images plays an important role in image guided interven- tion. However, the lack of clear boundary specifically at the apex and base, and huge variation of shape and texture between the images from different patients make the task very challenging. To overcome these problems, in this paper, we propose a deeply supervised convolutional neural network (CNN) utilizing the convolutional information to accurately segment the prostate from MR images. The proposed model can effectively detect the prostate region with additional deeply supervised layers compared with other approaches. Since some information will be abandoned after convolution, it is necessary to pass the features extracted from early stages to later stages. The experimental results show that significant segmentation accuracy improvement has been achieved by our proposed method compared to other reported approaches.
I. INTRODUCTION
CNNs have become powerful tools for medical image analysis, but prostate segmentation remains challenging because accurate pixel-level labels depend on weak boundaries and large patient-specific variation.
- CNNs learn hierarchical data representations and have achieved state-of-the-art performance in computer vision and medical image analysis.The paper links this progress to improved GPUs, large annotated datasets, and task-specific network designs.
- Medical image segmentation requires assigning a precise label to every pixel, making edge detection and boundary localization central concerns.Earlier approaches therefore focused on constructing effective features for identifying structure boundaries.
- Prostate segmentation is especially difficult because the apex and base often lack clear boundaries, while shape and texture vary substantially across patients.These challenges prevent direct application of standard fully convolutional networks to the task.
- Patch-to-pixel and patch-to-patch strategies reduce training and prediction efficiency, whereas fully convolutional networks enable image-to-image training on larger sample batches.The paper motivates an image-to-image approach while addressing prostate-specific segmentation difficulties.
- The proposed Deeply-Supervised CNN forwards early-stage features to later stages and adds supervised layers at each stage for accurate, fast prostate segmentation.The design targets information loss after convolution and gives hidden representations direct supervision.
II. METHODS
The baseline U-Net uses a contracting path for multiresolution feature extraction and an expanding path for progressively reconstructing feature maps.
- U-Net contains four stages on both its left and right network parts, with each stage processing feature maps at a different resolution.The left path performs feature extraction and downsampling, while the right path reconstructs representations.
- Each left-side stage uses two 3x3 convolutions followed by ReLU activations and 2x2 max pooling with stride 2 for downsampling.The number of feature channels doubles after each left-side stage.
B. Deeply-Supervised CNN
The proposed architecture combines deeper, parameter-efficient convolutions with residual feature transfer and deep supervision to preserve information and improve training for prostate segmentation.
- Network Architecture: 1x1 convolutions reduce dimensions and parameters while enabling a deeper network and improved feature representation.This addresses overfitting and computational-resource bottlenecks associated with deeper networks trained on limited medical-image datasets.
- Network Architecture: The architecture uses smaller 3x3 kernels and pooling operations to reduce parameters, ease computation, and mitigate overfitting.Stacked small kernels provide an effective receptive field comparable to a larger kernel.
- Deeply-Supervised CNN: Residual connections forward early-stage features to later stages, supplying information that convolution may lose before final prediction.The network contains a compression path with progressively reduced resolution and increased feature-channel counts.
- Deeply-Supervised CNN: Eight additional deeply supervised layers constrain training, make hidden features more semantically meaningful, and preserve gradients from early stages.These layers address weak semantic meaning in hidden representations and gradient propagation concerns in deep networks.
- Deeply-Supervised CNN: The additional supervised layers make residual information meaningful and improve the model’s convergence time.The paper presents this as an advantage over the original U-Net.
2) Formulation:
The formulation defines image inputs, binary ground-truth edge maps, and deep-supervision weights, then combines supervised-layer and final-output losses. Dice coefficient optimization addresses severe foreground–background imbalance in prostate segmentation.
- The training set pairs raw input images X_n with corresponding ground-truth binary edge maps Y_n.
- The network parameters are W, while deep-supervised layers use weights w={w_1,...,w_m}, with m=8 in this method.
- Dice coefficient is used as the objective because the prostate occupies a small image region and ordinary learning can become biased toward background.
- DSC compares automatic segmentation shape S_a with manual segmentation shape S_m.
- For binary images, DSC is computed over ground-truth pixels p_i and segmentation pixels q_i across N total pixels.
- The final objective combines deep-supervised losses with the final-output loss and is minimized using stochastic gradient descent.
A. Dataset
The dataset contains prostate MRI volumes from 81 patients, with non-prostate slices removed and four patients reserved for testing. Data augmentation expands the limited training data using geometric transformations.
- All images were acquired from 81 patients using a Philips 3T MRI scanner with an endorectal coil.
- Four patients totaling 64 images were randomly selected for testing, while the remaining patients were used for training.
- Translation, rotation, and zoom augmentation were applied to increase robustness and reduce overfitting with limited images.
B. Implementation Details
The network is trained end-to-end on fixed-size 512x512 prostate MRI images using Keras, with specified optimization settings and a batch size constrained by GPU memory.
- Training uses fixed-size 512x512 images and ground truths in an end-to-end Keras implementation.
- The learning rate starts at 0.001 and decreases through weight decay, with momentum set to 0.9.
- Batch size is 1 because of memory limitations, and experiments run on an 8 GB GTX1080 GPU using CUDA 8.0.
IV. RESULT AND DISCUSSION
The evaluation trains the proposed network and comparison models on the same dataset, using held-out patient images with radiologist annotations and Dice coefficient as the evaluation metric.
- The network was trained on 77 patients, while four patients with 64 images were held out before training for evaluation.
- The held-out prostate regions were manually predelineated by a radiologist and used as ground truth for automatic-segmentation evaluation.
- Table I presents a quantitative comparison between the proposed approach and other methods.
- Dice coefficient was selected as the evaluation method.
- U-Net and fully convolutional networks were trained and tested on the same dataset for validation against the proposed method.
1) Qualitative Comparison:
Qualitative comparisons show that deeply supervised CNN better preserves prostate boundaries than FCN and U-Net on challenging MR images.
- Challenging samples contain fuzzy boundaries, inhomogeneous intensities, and similar prostate and nonprostate contrast.
- FCN detects part of the prostate but assigns boundary pixels inaccurately.
- U-Net improves pixel labeling but mislabels nonprostate regions and poorly captures boundaries.
- Deeply supervised CNN detects fuzzy boundaries with more continuous and smoother segmentation contours.The authors attribute these improvements to additional supervised layers capturing prostate texture and boundary information.
2) Quantitative Comparison:
Quantitative evaluation reports higher DSC values for the deeply supervised CNN than for U-Net and FCN, with additional evidence addressing fuzzy boundaries and intensity variation.
- 0.885 average DSC is reported for the proposed model, compared with 0.865 for U-Net and 0.759 for FCN.The proposed method also has the highest average, median, and maximum DSC values.
- The proposed method is reported to improve prostate segmentation relative to U-Net and FCN.The authors attribute the improvement to a deeper network and additional supervised layers that constrain training.
- The method addresses fuzzy boundaries and inhomogeneous pixel-intensity distributions inside and outside the prostate.
3) Discussion:
The discussion links deep supervision to reduced prostate-information loss and improved gradient flow, while the conclusion combines residual information, compact depth, and Dice optimization.
- Deep supervision provides gradients to early network stages, helping address gradient diffusion in a deep architecture.
- Early supervised layers retain more information, while later layers produce representations closer to the ground truth.Different supervised layers are reported to detect different textures.
- The proposed CNN uses residual information and 1x1 convolution to deepen the network without simultaneously increasing parameter count relative to traditional U-Net.
- Dice coefficient optimization addresses the imbalance between prostate foreground pixels and background pixels.The conclusion reports that the proposed network improves segmentation performance.