A New Ensemble Learning Framework for 3D Biomedical Image Segmentation
Authors: Hao Zheng, Yizhe Zhang, Lin Yang, Peixian Liang, Zhuo Zhao, Chaoli Wang, Danny Z. Chen5909-5916
AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on two public datasets (the HVSMR 2016 Challenge dataset and the mouse piriform cortex dataset) show that our approach is effective under fully-supervised, semi-supervised, and transductive settings, and attains superior performance over state-of-the-art image segmentation methods. |
| Researcher Affiliation | Academia | Hao Zheng, Yizhe Zhang, Lin Yang, Peixian Liang, Zhuo Zhao, Chaoli Wang, Danny Z. Chen Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA {hzheng3, yzhang29, lyang5, pliang, zzhao3, cwang11, dchen}@nd.edu |
| Pseudocode | Yes | Algorithm 1: Random-fit Input: (xi, PLi = {f1(xi), f2(xi), . . . , fm(xi)}, S(PLi)) , i = 1, 2, . . . , n; Output: A trained meta-learner H; initialize a meta-learner H with random weights; mini-batch = ; while stopping condition not met do for k = 1 to batch-size do p = rand-int(1, n); q = rand-int(1, m); add training sample {(xp, S(PLp)), fq(xp)} to the mini-batch; update H using training samples in the mini-batch with forward and backward propagation; mini-batch = ; |
| Open Source Code | Yes | Code will be made publicly available at https://github.com/Hao Zheng94/Ensemble. |
| Open Datasets | Yes | We evaluate our approach using two public datasets: (1) the HVSMR 2016 Challenge dataset (Pace et al. 2015) and (2) the mouse piriform cortex dataset (Lee et al. 2015). |
| Dataset Splits | Yes | HVSMR 2016. The objective of the HVSMR 2016 Challenge (Pace et al. 2015) is to segment the myocardium and great vessel (blood pool) in cardiovascular magnetic resonance (MR) images. 10 3D MR images and their corresponding ground truth annotation are provided by the challenge organizers as training data. The test data, consisting of another 10 3D MR images, are publicly available, yet their ground truths are kept secret for fair comparison. The results are evaluated using three criteria: (1) Dice coefficient, (2) average surface distance (ADB), and (3) symmetric Hausdorff distance. Finally, a score S, computed as S = P 4ADB 1 30Hausdorff), is used to reflect the overall accuracy of the results and for ranking. Mouse piriform cortex. Our approach is also evaluated on the mouse piriform cortex dataset (Lee et al. 2015) for neuron boundary segmentation in serial section EM images. This dataset contains 4 stacks of 3D EM images. Following the previous practice (Lee et al. 2015; Shen et al. 2017), the 2nd, 3rd, and 4th stacks are used for model training, and the 1st stack is used for testing. Also, as in (Lee et al. 2015; Shen et al. 2017), the results are evaluated using the Rand Fscore (the harmonic mean of the Rand merge score and the Rand split score). |
| Hardware Specification | No | The paper does not explicitly specify the hardware used (e.g., specific GPU or CPU models). |
| Software Dependencies | Yes | All our networks are implemented using Tensor Flow (Abadi et al. 2016). All our networks are trained using Adam (Kingma and Ba 2015) with β1 = 0.9, β2 = 0.999, and ϵ = 1e-10. |
| Experiment Setup | Yes | The initial learning rates are all set as 5e-4. Our 2D base-learners reduce the learning rates to 5e-5 after 10k iterations; our 3D base-learner and meta learner adopt the poly learning rate policy (Yu et al. 2017) with the power variable equal to 0.9 and the max iteration number equal to 40k. To leverage the limited training data, standard data augmentation techniques (i.e., random rotation with 90, 180, and 270 degrees, as well as image flipping along the axial planes) are employed to augment the training data. For the HVSMR 2016 Challenge dataset, due to large intensity variance among different images, all the cardiac images are normalized to have zero mean and unit variance. We also employ spatial resampling to 1mm isotropically. |