Source-linked AI summary
V-Net: Fully Convolutional Neural Networks for Volumetric Medical Image Segmentation
Fausto Milletari, Nassir Navab, Seyed-Ahmad Ahmadi
TL;DR
Prostate MRI segmentation requires volumetric methods that handle challenging appearance variation and imaging artefacts. V-Net performs whole-volume segmentation with a 3D fully convolutional network and Dice-based objective, achieving fast and accurate results on test volumes.
Problem
Prostate MRI segmentation is challenging because appearance varies across scans and volumes contain artefacts, yet accurate boundaries matter for diagnosis and treatment planning.
Method
V-Net uses an end-to-end volumetric fully convolutional network with Dice-based loss to segment entire prostate MRI volumes.
Results
V-Net produces fast and accurate prostate MRI segmentations and directly compares favorably with methods evaluated on the same test data.
Takeaways & Limitations
The Dice loss supports binary segmentation under strong foreground–background imbalance without sample re-weighting.
Takeaways & Limitations
The approach required data augmentation because annotated medical volumes are difficult and costly to obtain.
Abstract
from arXiv · showhide
Convolutional Neural Networks (CNNs) have been recently employed to solve problems from both the computer vision and medical image analysis fields. Despite their popularity, most approaches are only able to process 2D images while most medical data used in clinical practice consists of 3D volumes. In this work we propose an approach to 3D image segmentation based on a volumetric, fully convolutional, neural network. Our CNN is trained end-to-end on MRI volumes depicting prostate, and learns to predict segmentation for the whole volume at once. We introduce a novel objective function, that we optimise during training, based on Dice coefficient. In this way we can deal with situations where there is a strong imbalance between the number of foreground and background voxels. To cope with the limited number of annotated volumes available for training, we augment the data applying random non-linear transformations and histogram matching. We show in our experimental evaluation that our approach achieves good performances on challenging test data while requiring only a fraction of the processing time needed by other previous methods.
1 Introduction and Related Work
The paper addresses challenging prostate MRI segmentation by processing entire 3D volumes with an end-to-end volumetric fully convolutional network. Its architecture combines compression and decompression paths, residual stages, skip connections, and a Dice-based objective to preserve detail while capturing whole-volume context.
- Motivation: 3D volumetric segmentation is important for medical imagery because diagnostic and interventional data often consists of full volumes.Automatic delineation supports visual augmentation, computer-assisted diagnosis, interventions, and quantitative image indices.
- Motivation: Prostate MRI segmentation is challenging because scans exhibit anatomical deformations, intensity variations, field-inhomogeneity artefacts, and distortions.The work therefore focuses on segmenting prostate MRI volumes while accounting for whole-volume content.
- Contribution: The proposed network processes MRI volumes end-to-end with volumetric convolutions rather than slice-wise processing and optimizes a novel Dice-coefficient objective.The authors report fast and accurate prostate MRI segmentation results.
- Network architecture: The architecture uses a compression path with multi-resolution residual stages and a decompression path that reconstructs the original spatial size for two-channel volumetric segmentation.Downsampling uses strided convolutions, while de-convolutions expand lower-resolution feature maps before further convolutional processing.
- Network architecture: Skip connections forward early-stage features to the decoder, restoring fine-grained information and improving final contour quality and convergence time.These horizontal connections address detail lost during compression.
- Network architecture: The network’s innermost features capture the whole input volume, enabling deepest layers to perceive poorly visible anatomy at once.The reported receptive-field analysis supports whole-volume spatial context in the deepest portion of the CNN.
3 Dice loss layer
The Dice loss layer operates on soft-max voxel probabilities to address severe foreground–background imbalance in medical volumes. Its differentiable formulation avoids class-reweighting and was experimentally better than multinomial logistic loss with sample re-weighting.
- Dice loss layer: Soft-max outputs foreground and background probabilities for every voxel, while small target anatomy can trap training in unfavorable loss minima.The network produces two volumes at the input resolution before soft-max processing.
- Dice loss layer: The Dice formulation sums over predicted and ground-truth binary volumes and can be differentiated to obtain voxel-wise prediction gradients.The gradient is computed with respect to the j-th prediction voxel.
- Dice loss layer: Dice optimization avoids assigning class weights and experimentally outperformed multinomial logistic loss trained with sample re-weighting.The comparison is reported for the same network and referenced in Figure 6.
4 Results
V-Net was trained and evaluated on PROMISE 2012 prostate MRI volumes spanning clinical and acquisition variability. Performance was assessed with segmentation metrics and challenge scores, using volumetric-convolution experiments that trained for about 48 hours and segmented unseen volumes in about one second.
- Experimental setup: The method was tested on 30 prostate MRI volumes with secret ground-truth annotations obtained from the challenge organisers.The test set represented clinical variability encountered in real prostate scans.
- Evaluation: Performance was evaluated using Dice coefficient, Hausdorff distance, and the PROMISE 2012 challenge score.Results were reported in Table 2 and Fig. 5, alongside qualitative results on the dataset.
- Implementation and runtime: The implementation used volumetric convolutions in a custom Caffe framework with CuDNN v3 on a workstation with an NVIDIA GTX 1080 GPU.The workstation also included 64 GB of memory and an Intel Core i7-5820K CPU at 3.30 GHz.
- Implementation and runtime: Training lasted about 48 hours, or approximately 30K iterations, while segmenting an unseen volume took about 1 second.These timings were measured for the reported implementation and workstation configuration.
5 Conclusion
The paper presents a volumetric convolutional neural network for fast and accurate prostate MRI volume segmentation. It introduces a Dice-based objective that handles strong foreground–background imbalance without sample re-weighting.
- 5 Conclusion: The proposed volumetric convolutional neural network segments prostate MRI volumes in a fast and accurate manner.The method performs segmentation directly on MRI prostate volumes.
- 5 Conclusion: The training objective is based on the Dice overlap coefficient between predicted segmentations and ground-truth annotations.The objective is implemented as a Dice loss layer optimized during training.
- 5 Conclusion: The Dice loss handles strong foreground–background imbalance without requiring sample re-weighting.This property is presented for binary segmentation.