Biomedical Image Datasets Research Articles

Biomedical image classification using seagull optimization with deep learning for colon and lung cancer diagnosis

<p>Traditional health care relies on biomedical image categorization to identify and treat various medical conditions. In machine learning and medical imaging, biomedical image classification for colon and lung cancer diagnosis is significant. The work focuses on building novel models and algorithms to accurately detect and categorize tumorous lesions using computer tomography (CT) scans and histopathology slides. These systems use image processing, deep learning (DL), and convolutional neural networks (CNN) to assist medical professionals diagnose cancer sooner and improve patient outcomes. Biomedical image classification using seagull optimization with deep learning (BIC-SGODL) addresses colon and lung cancer diagnosis. The BIC-SGODL method improves cancer diagnosis using hyperparameter optimized DL model. BIC-SGODL utilizes DenseNet to learn complicated features. The convolutional long short-term memory (CLSTM) standard captures spatiotemporal information in sequential picture data. Finally, the SGO method adjusts hyperparameters to improve model performance and generalization. BIC-SGODL performs well with biomedical image dataset simulations. Thus, medical picture cancer diagnosis may be automated using BIC-SGODL.</p>

Towards laryngeal cancer diagnosis using Dandelion Optimizer Algorithm with ensemble learning on biomedical throat region images

Laryngeal cancer exhibits a notable global health burden, with later-stage detection contributing to a low mortality rate. Laryngeal cancer diagnosis on throat region images is a pivotal application of computer vision (CV) and medical image diagnoses in the medical sector. It includes detecting and analysing abnormal or cancerous tissue from the larynx, an integral part of the vocal and respiratory systems. The computer-aided system makes use of artificial intelligence (AI) through deep learning (DL) and machine learning (ML) models, including convolution neural networks (CNN), for automated disease diagnoses and detection. Various DL and ML approaches are executed to categorize the extraction feature as healthy and cancerous tissues. This article introduces an automated Laryngeal Cancer Diagnosis using the Dandelion Optimizer Algorithm with Ensemble Learning (LCD-DOAEL) method on Biomedical Throat Region Image. The LCD-DOAEL method aims to investigate the images of the throat region for the presence of laryngeal cancer. In the LCD-DOAEL method, the Gaussian filtering (GF) approach is applied to eliminate the noise in the biomedical images. Besides, the complex and intrinsic feature patterns can be extracted by the MobileNetv2 model. Meanwhile, the DOA model carries out the hyperparameter selection of MobileNetV2 architecture. Finally, the ensemble of three classifiers such as bidirectional long short-term memory (BiLSTM), regularized extreme learning machine (ELM), and backpropagation neural network (BPNN) models, are utilized for the classification process. A comprehensive set of simulations is conducted on the biomedical image dataset to highlight the efficient performance of the LCD-DOAEL technique. The comparison analysis of the LCD-DOAEL method exhibited a superior accuracy outcome of 97.54% over other existing techniques.

MRAU-net: Multi-scale residual attention U-shaped network for medical image segmentation

Medical image segmentation is a pivotal technology that assists doctors in making correct diagnoses and provides reliable evidence for pathological research. To address the insufficient attention to medical image details and the weak generalization ability of convolutional neural networks, we propose a novel multi-scale residual attention U-shaped network (MRAU-Net) for medical image segmentation. MRAU-Net employs a channel-wise compression strategy to construct the squeezed double convolution in the encoder, significantly reducing the network parameters while maintaining segmentation performance. In the decoder, a multi-scale residual attention block is designed to learn rich context information with multiple receptive fields, which effectively focuses on the target area and compensates for the lost information during the upsampling process through skip connections, enhancing the network's feature extraction capability. Experimental results on four biomedical image datasets show that our MRAU-Net performs better than the existing methods in terms of segmentation accuracy and generalization ability. mDices of MRAU-Net on JSUAHCerebellum, Kvasir-SEG, ISIC 2018 Challenge, and 2018 Data Science Bowl are 92.81 %, 88.64 %, 91.88 %, and 93.14 %, respectively, which are 7.61 %, 7.48 %, 6.34 %, and 2.34 % higher than that of the original U-Net.

RCI-Seg: Robust click-based interactive segmentation framework with deep reinforcement learning for biomedical images

Recently, interactive segmentation models have achieved remarkable success in the field of biomedical images. However, these models rely on the accurate and high-quality interaction information provided by users, otherwise the segmentation performance will be seriously affected. This problem is more severe in multi-target biomedical images, which means that an image contains multiple targets of interest. It is extremely challenging for users to always maintain high-quality interactions. In this paper, we propose a novel two-stage segmentation model with robust interaction points for biomedical images. In the first stage, we implement robust interaction points based on user initial interaction points and the deep reinforcement learning (DRL) model. Specifically, we build a reinforcement learning environment to simulate the movement of interaction points with agents, and obtain improved interaction points (clue points) that are beneficial for segmentation. In the second stage, we use a convolutional neural network (CNN) model to achieve segmentation by combining clue points and biomedical image. We validate the performance of our approach on five public biomedical image datasets. The experimental results show that the proposed approach outperforms several SOTA methods in multiple metrics.

A knowledge-based learning framework for self-supervised pre-training towards enhanced recognition of biomedical microscopy images

Self-supervised pre-training has become the priory choice to establish reliable neural networks for automated recognition of massive biomedical microscopy images, which are routinely annotation-free, without semantics, and without guarantee of quality. Note that this paradigm is still at its infancy and limited by closely related open issues: (1) how to learn robust representations in an unsupervised manner from unlabeled biomedical microscopy images of low diversity in samples? and (2) how to obtain the most significant representations demanded by a high-quality segmentation? Aiming at these issues, this study proposes a knowledge-based learning framework (TOWER) towards enhanced recognition of biomedical microscopy images, which works in three phases by synergizing contrastive learning and generative learning methods: (1) Sample Space Diversification: Reconstructive proxy tasks have been enabled to embed a priori knowledge with context highlighted to diversify the expanded sample space; (2) Enhanced Representation Learning: Informative noise-contrastive estimation loss regularizes the encoder to enhance representation learning of annotation-free images; (3) Correlated Optimization: Optimization operations in pre-training the encoder and the decoder have been correlated via image restoration from proxy tasks, targeting the need for semantic segmentation. Experiments have been conducted on public datasets of biomedical microscopy images against the state-of-the-art counterparts (e.g., SimCLR and BYOL), and results demonstrate that: TOWER statistically excels in all self-supervised methods, achieving a Dice improvement of 1.38 percentage points over SimCLR. TOWER also has potential in multi-modality medical image analysis and enables label-efficient semi-supervised learning, e.g., reducing the annotation cost by up to 99% in pathological classification.

BASIN: A Semi-automatic Workflow, with Machine Learning Segmentation, for Objective Statistical Analysis of Biomedical and Biofilm Image Datasets

Micrograph comparison remains useful in bioscience. This technology provides researchers with a quick snapshot of experimental conditions. But sometimes a two- condition comparison relies on researchers’ eyes to draw conclusions. Our Bioimage Analysis, Statistic, and Comparison (BASIN) software provides an objective and reproducible comparison leveraging inferential statistics to bridge image data with other modalities. Users have access to machine learning-based object segmentation. BASIN provides several data points such as images’ object counts, intensities, and areas. Hypothesis testing may also be performed. To improve BASIN’s accessibility, we implemented it using R Shiny and provided both an online and offline version. We used BASIN to process 498 image pairs involving five bioscience topics. Our framework supported either direct claims or extrapolations 57% of the time. Analysis results were manually curated to determine BASIN’s accuracy which was shown to be 78%. Additionally, each BASIN version’s initial release shows an average 82% FAIR compliance score.

GAPCNN with HyPar: Global Average Pooling convolutional neural network with novel NNLU activation function and HYBRID parallelism.

With the increasing demand for deep learning in the last few years, CNNs have been widely used in many applications and have gained interest in classification, regression, and image recognition tasks. The training of these deep neural networks is compute-intensive and takes days or even weeks to train the model from scratch. The compute-intensive nature of these deep neural networks sometimes limits the practical implementation of CNNs in real-time applications. Therefore, the computational speedup in these networks is of utmost importance, which generates interest in CNN training acceleration. Much research is going on to meet the computational requirement and make it feasible for real-time applications. Because of its simplicity, data parallelism is used primarily, but it performs badly sometimes. In most cases, researchers prefer model parallelism to data parallelism, but it is not always the best choice. Therefore, in this study, we implement a hybrid of both data and model parallelism to improve the computational speed without compromising accuracy. There is only a 1.5% accuracy drop in our proposed study with an increased speed up of 3.62X. Also, a novel activation function Normalized Non-linear Activation Unit NNLU is proposed to introduce non-linearity in the model. The activation unit is non-saturated and helps avoid the model's over-fitting. The activation unit is free from the vanishing gradient problem. Also, the fully connected layer in the proposed CNN model is replaced by the Global Average Pooling layers (GAP) to enhance the model's accuracy and computational performance. When tested on a bio-medical image dataset, the model achieves an accuracy of 98.89% and requires a training time of only 1 s. The model categorizes medical images into different categories of glioma, meningioma, and pituitary tumor. The model is compared with existing state-of-art techniques, and it is observed that the proposed model outperforms others in classification accuracy and computational speed. Also, results are observed for different optimizers', different learning rates, and various epoch numbers.

Deep Learning Assessment for Mining Important Medical Image Features of Various Modalities.

Deep learning (DL) is a well-established pipeline for feature extraction in medical and nonmedical imaging tasks, such as object detection, segmentation, and classification. However, DL faces the issue of explainability, which prohibits reliable utilisation in everyday clinical practice. This study evaluates DL methods for their efficiency in revealing and suggesting potential image biomarkers. Eleven biomedical image datasets of various modalities are utilised, including SPECT, CT, photographs, microscopy, and X-ray. Seven state-of-the-art CNNs are employed and tuned to perform image classification in tasks. The main conclusion of the research is that DL reveals potential biomarkers in several cases, especially when the models are trained from scratch in domains where low-level features such as shapes and edges are not enough to make decisions. Furthermore, in some cases, device acquisition variations slightly affect the performance of DL models.

Addressing the Intra-class Mode Collapse Problem using Adaptive Input Image Normalization in GAN-based X-ray Images.

Biomedical image datasets can be imbalanced due to the rarity of targeted diseases. Generative Adversarial Networks play a key role in addressing this imbalance by enabling the generation of synthetic images to augment datasets. It is important to generate synthetic images that incorporate a diverse range of features to accurately represent the distribution of features present in the training imagery. Furthermore, the absence of diverse features in synthetic images can degrade the performance of machine learning classifiers. The mode collapse problem impacts Generative Adversarial Networks' capacity to generate diversified images. Mode collapse comes in two varieties: intra-class and inter-class. In this paper, the intra-class mode collapse problem is investigated, and its subsequent impact on the diversity of synthetic X-ray images is evaluated. This work contributes an empirical demonstration of the benefits of integrating the adaptive input-image normalization for the Deep Convolutional GAN to alleviate the intra-class mode collapse problem. Results demonstrate that the DCGAN with adaptive input-image normalization outperforms DCGAN with un-normalized X-ray images as evident by the superior diversity scores.

Image Augmentation Techniques for Mammogram Analysis.

Research in the medical imaging field using deep learning approaches has become progressively contingent. Scientific findings reveal that supervised deep learning methods’ performance heavily depends on training set size, which expert radiologists must manually annotate. The latter is quite a tiring and time-consuming task. Therefore, most of the freely accessible biomedical image datasets are small-sized. Furthermore, it is challenging to have big-sized medical image datasets due to privacy and legal issues. Consequently, not a small number of supervised deep learning models are prone to overfitting and cannot produce generalized output. One of the most popular methods to mitigate the issue above goes under the name of data augmentation. This technique helps increase training set size by utilizing various transformations and has been publicized to improve the model performance when tested on new data. This article surveyed different data augmentation techniques employed on mammogram images. The article aims to provide insights into basic and deep learning-based augmentation techniques.

A method for balancing a multi-labeled biomedical dataset

In this paper, we propose a data balancing method for multi-label biomedical data. The method can be applied in the case of semantic segmentation problems for balancing the corresponding image data. The proposed method performs oversampling of instances of minority classes in a way that increases the frequencies of appearance (a ratio of number of samples, containing this class, over the total number of samples in the dataset) of minority classes in the data, thereby reducing the class imbalance. The effectiveness of the proposed method is shown experimentally by applying it to two highly unbalanced biomedical image datasets. A convolutional neural network (CNN) was trained on several versions of those datasets: one balanced with the proposed method, another balanced with manual oversampling and an unbalanced version. The results of the experiments validate the effectiveness of the proposed method, proving that it allows the influence of class imbalance on the learning algorithm to be reduced, thus improving its original classification results for most of the classes. Apart from biomedical image data, the proposed method was applied to several common multi-label datasets. Inherently, the proposed method does not make any assumptions about the underlying structure of the data to be balanced; therefore, it can be applied to all types of data (vectors, images, etc.) that can be described in a multi-label framework. It also can be used in conjunction with any learning algorithm that is suitable for multi-label data. To illustrate its wider applicability, a series of experiments was conducted using seven common multi-label datasets. An experimental comparison to existing multi-label data balancing approaches is provided, as well. The experimental results show that the proposed method presents a competitive alternative to existing approaches.

Hierarchical confounder discovery in the experiment-machine learning cycle

SummaryThe promise of machine learning (ML) to extract insights from high-dimensional datasets is tempered by confounding variables. It behooves scientists to determine if a model has extracted the desired information or instead fallen prey to bias. Due to features of natural phenomena and experimental design constraints, bioscience datasets are often organized in nested hierarchies that obfuscate the origins of confounding effects and render confounder amelioration methods ineffective. We propose a non-parametric statistical method called the rank-to-group (RTG) score that identifies hierarchical confounder effects in raw data and ML-derived embeddings. We show that RTG scores correctly assign the effects of hierarchical confounders when linear methods fail. In a public biomedical image dataset, we discover unreported effects of experimental design. We then use RTG scores to discover crossmodal correlated variability in a multi-phenotypic biological dataset. This approach should be generally useful in experiment-analysis cycles and to ensure confounder robustness in ML models.

Evaluation of the benchmark datasets for testing the efficacy of deep convolutional neural networks

In the past decade, deep neural networks, and specifically convolutional neural networks (CNNs), have been becoming a primary tool in the field of biomedical image analysis, and are used intensively in other fields such as object or face recognition. CNNs have a clear advantage in their ability to provide superior performance, yet without the requirement to fully understand the image elements that reflect the biomedical problem at hand, and without designing specific algorithms for that task. The availability of easy-to-use libraries and their non-parametric nature make CNN the most common solution to problems that require automatic biomedical image analysis. But while CNNs have many advantages, they also have certain downsides. The features determined by CNNs are complex and unintuitive, and therefore CNNs often work as a “Black Box”. Additionally, CNNs learn from any piece of information in the pixel data that can provide a discriminative signal, making it more difficult to control what the CNN actually learns. Here we follow common practices to test whether CNNs can classify biomedical image datasets, but instead of using the entire image we use merely parts of the images that do not have biomedical content. The experiments show that CNNs can provide high classification accuracy even when they are trained with datasets that do not contain any biomedical information, or can be systematically biased by irrelevant information in the image data. The presence of such consistent irrelevant data is difficult to identify, and can therefore lead to biased experimental results. Possible solutions to this downside of CNNs can be control experiments, as well as other protective practices to validate the results and avoid biased conclusions based on CNN-generated annotations.

Combined Transfer Learning and Test-Time Augmentation Improves Convolutional Neural Network-Based Semantic Segmentation of Prostate Cancer from Multi-Parametric MR Images

PurposeMultiparametric MRI (mp-MRI) is a widely used tool for diagnosing and staging prostate cancer. The purpose of this study was to evaluate whether transfer learning, unsupervised pre-training and test-time augmentation significantly improved the performance of a convolutional neural network (CNN) for pixel-by-pixel prediction of cancer vs. non-cancer using mp-MRI datasets. Methods154 subjects undergoing mp-MRI were prospectively recruited, 16 of whom subsequently underwent radical prostatectomy. Logistic regression, random forest and CNN models were trained on mp-MRI data using histopathology as the gold standard. Transfer learning, unsupervised pre-training and test-time augmentation were used to boost CNN performance. Models were evaluated using Dice score and area under the receiver operating curve (AUROC) with leave-one-subject-out cross validation. Permutation feature importance testing was performed to evaluate the relative value of each MR contrast to CNN model performance. Statistical significance (p<0.05) was determined using the paired Wilcoxon signed rank test with Benjamini-Hochberg correction for multiple comparisons. ResultsBaseline CNN outperformed logistic regression and random forest models. Transfer learning and unsupervised pre-training did not significantly improve CNN performance over baseline; however, test-time augmentation resulted in significantly higher Dice scores over both baseline CNN and CNN plus either of transfer learning or unsupervised pre-training. The best performing model was CNN with transfer learning and test-time augmentation (Dice score of 0.59 and AUROC of 0.93). The most important contrast was apparent diffusion coefficient (ADC), followed by Ktrans and T2, although each contributed significantly to classifier performance. ConclusionsThe addition of transfer learning and test-time augmentation resulted in significant improvement in CNN segmentation performance in a small set of prostate cancer mp-MRI data. Results suggest that these techniques may be more broadly useful for the optimization of deep learning algorithms applied to the problem of semantic segmentation in biomedical image datasets. However, further work is needed to improve the generalizability of the specific model presented herein.

CNL-UNet: A novel lightweight deep learning architecture for multimodal biomedical image segmentation with false output suppression

Automatic biomedical image segmentation plays an important role in speeding up disease detection and diagnosis. The rapid development of Deep Learning has shown ground-breaking improvements in this context. However, state-of-the-art networks like U-Net and SegNet often have poor performance on challenging domains. Most of the recent works were domains specific and computationally expensive. This paper proposes a novel lightweight architecture named CNL-UNet for 2D multimodal Biomedical Image Segmentation. The proposed CNL-UNet has a pre-trained encoder enriched with transfer learning techniques to learn sufficiently from the small amount of data. It has modified skip connections to reduce semantic gaps between the corresponding level of the encoder-decoder layer. Furthermore, the proposed architecture is enhanced with a novel Classifier and Localizer (CNL) module. This module provides us with additional classification and localization information with greater accuracy. Fusing this information with the segmentation output, the CNL-UNet can suppress false positives and false negatives. The proposed architcture has comparatively fewer parameters (11.5M) than U-Net (31M), SegNet (29M), and most of the recent works. Thus it is a lightweight architecture and also less prone to overfit. Besides, in the case of simple datasets, the pruned version of the CNL-UNet can be used. We evaluated our proposed architecture on multimodal biomedical image datasets, namely Chest X-ray, Dermoscopy, Microscopy, Ultrasound, and MRI images. The results demonstrate the superior performance of our proposed architecture over most of the existing networks. We have shown that our model can learn quickly, segment precisely, and automatically suppress falsely classified outputs.

Convolutional neuronal networks combined with X-ray phase-contrast imaging for a fast and observer-independent discrimination of cartilage and liver diseases stages

We applied transfer learning using Convolutional Neuronal Networks to high resolution X-ray phase contrast computed tomography datasets and tested the potential of the systems to accurately classify Computed Tomography images of different stages of two diseases, i.e. osteoarthritis and liver fibrosis. The purpose is to identify a time-effective and observer-independent methodology to identify pathological conditions. Propagation-based X-ray phase contrast imaging WAS used with polychromatic X-rays to obtain a 3D visualization of 4 human cartilage plugs and 6 rat liver samples with a voxel size of 0.7 × 0.7 × 0.7 µm3 and 2.2 × 2.2 × 2.2 µm3, respectively. Images with a size of 224 × 224 pixels are used to train three pre-trained convolutional neuronal networks for data classification, which are the VGG16, the Inception V3, and the Xception networks. We evaluated the performance of the three systems in terms of classification accuracy and studied the effect of the variation of the number of inputs, training images and of iterations. The VGG16 network provides the highest classification accuracy when the training and the validation-test of the network are performed using data from the same samples for both the cartilage (99.8%) and the liver (95.5%) datasets. The Inception V3 and Xception networks achieve an accuracy of 84.7% (43.1%) and of 72.6% (53.7%), respectively, for the cartilage (liver) images. By using data from different samples for the training and validation-test processes, the Xception network provided the highest test accuracy for the cartilage dataset (75.7%), while for the liver dataset the VGG16 network gave the best results (75.4%). By using convolutional neuronal networks we show that it is possible to classify large datasets of biomedical images in less than 25 min on a 8 CPU processor machine providing a precise, robust, fast and observer-independent method for the discrimination/classification of different stages of osteoarthritis and liver diseases.

Online variational learning of finite inverted Beta‐Liouville mixture model for biomedical analysis

AbstractImage segmentation is widely applied for biomedical image analysis. However, segmentation of medical images is challenging due to many image modalities, such as, CT, X‐ray, MRI, microscopy among others. An additional challenge to this is the high variability, inconsistent regions with missing edges, absence of texture contrast, and high noise in the background of biomedical images. Thus, many segmentation approaches have been investigated to address these issues and to transform medical images into meaningful information. During the past decade, finite mixture models have been revealed to be one of the most flexible and popular approaches in data clustering. In this article, we propose a statistical framework for online variational learning of finite inverted Beta‐Liouville mixture model for clustering medical images. The online variational learning framework is used to estimate the parameters and the number of mixture components simultaneously, thus decreasing the computational complexity of the model. To this end, we evaluated our proposed algorithm on five different biomedical image data sets including optic disc detection and localization in diabetic retinopathy, digital imaging in melanoma lesion detection and segmentation, brain tumor detection, colon cancer detection and computer aid detection (CAD) of Malaria. Furthermore, we compared the proposed algorithm with three other popular algorithms. In our results, we analyze that the proposed online variational learning of finite IBL mixture model algorithm performs accurately on multiple modalities of medical images. It detects the disease patterns with high confidence. Computational and statistical approaches like the one presented in this article hold a significant impact on medical image analysis and interpretation in both clinical applications and scientific research. We believe that the proposed algorithm has the capacity to address multi modal biomedical image data sets and can be further applied by researchers to analyze correct disease patterns.

Color Enhancement Strategies for 3D Printing of X-ray Computed Tomography Bone Data for Advanced Anatomy Teaching Models

Three-dimensional (3D) printed anatomical models are valuable visual aids that are widely used in clinical and academic settings to teach complex anatomy. Procedures for converting human biomedical image datasets, like X-ray computed tomography (CT), to prinTable 3D files were explored, allowing easy reproduction of highly accurate models; however, these largely remain monochrome. While multi-color 3D printing is available in two accessible modalities (binder-jetting and poly-jet/multi-jet systems), studies embracing the viability of these technologies in the production of anatomical teaching models are relatively sparse, especially for sub-structures within a segmentation of homogeneous tissue density. Here, we outline a strategy to manually highlight anatomical subregions of a given structure and multi-color 3D print the resultant models in a cost-effective manner. Readily available high-resolution 3D reconstructed models are accessible to the public in online libraries. From these databases, four representative files (of a femur, lumbar vertebra, scapula, and innominate bone) were selected and digitally color enhanced with one of two strategies (painting or splitting) guided by Feneis and Dauber’s Pocket Atlas of Human Anatomy. Resulting models were created via 3D printing with binder-jet and/or poly-jet machines with important features, such as muscle origin and insertion points, highlighted using multiple colors. The resulting multi-color, physical models are promising teaching tools that will enhance the anatomical learning experience.

Bayesian inference framework for bounded generalized Gaussian‐based mixture model and its application to biomedical images classification

AbstractBiomedical image classification problem has attracted a lot of attention in medical engineering community and medicine applications. Accurate and automatic classification (eg, normal/abnormal or malignant/benign) has a variety of applications such as automatic decision making and is known to be very challenging. In this research, we address this problem by investigating the effectiveness of Bayesian inference methods for statistical bounded mixture models. Indeed, a novel approach termed as Bayesian learning for bounded generalized Gaussian mixture models is developed. The consideration of bounded mixture models is encouraged by their capability to take into account the nature of the data that is compactly supported. Furthermore, the consideration of Bayesian inference is more attractive compared to frequentist reasoning. In this work, we address main issues related to accurate data classification such as the effective estimation of the model's parameters and the selection of the optimal model complexity. Moreover, the problem of over‐ or under‐fitting is treated by taking into account the uncertainty through introducing prior information about the model's parameters. A comparative study between different Gaussian‐based models is also performed to evaluate the performance of the proposed framework. Experiments have been conducted on challenging biomedical image datasets that involve retinal images for diabetic retinopathy detection and mammograms for breast cancer detection. Obtained results are encouraging and show the benefits of our Bayesian framework.

A New Ensemble Learning Framework for 3D Biomedical Image Segmentation

3D image segmentation plays an important role in biomedical image analysis. Many 2D and 3D deep learning models have achieved state-of-the-art segmentation performance on 3D biomedical image datasets. Yet, 2D and 3D models have their own strengths and weaknesses, and by unifying them together, one may be able to achieve more accurate results. In this paper, we propose a new ensemble learning framework for 3D biomedical image segmentation that combines the merits of 2D and 3D models. First, we develop a fully convolutional network based meta-learner to learn how to improve the results from 2D and 3D models (base-learners). Then, to minimize over-fitting for our sophisticated meta-learner, we devise a new training method that uses the results of the baselearners as multiple versions of “ground truths”. Furthermore, since our new meta-learner training scheme does not depend on manual annotation, it can utilize abundant unlabeled 3D image data to further improve the model. 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, semisupervised, and transductive settings, and attains superior performance over state-of-the-art image segmentation methods.