Michael Fulham

Mingyuan Meng, Lei Bi, Michael Fulham et al.

Image registration is a fundamental requirement for medical image analysis. Deep registration methods based on deep learning have been widely recognized for their capabilities to perform fast end-to-end registration. Many deep registration methods achieved state-of-the-art performance by performing coarse-to-fine registration, where multiple registration steps were iterated with cascaded networks. Recently, Non-Iterative Coarse-to-finE (NICE) registration methods have been proposed to perform coarse-to-fine registration in a single network and showed advantages in both registration accuracy and runtime. However, existing NICE registration methods mainly focus on deformable registration, while affine registration, a common prerequisite, is still reliant on time-consuming traditional optimization-based methods or extra affine registration networks. In addition, existing NICE registration methods are limited by the intrinsic locality of convolution operations. Transformers may address this limitation for their capabilities to capture long-range dependency, but the benefits of using transformers for NICE registration have not been explored. In this study, we propose a Non-Iterative Coarse-to-finE Transformer network (NICE-Trans) for image registration. Our NICE-Trans is the first deep registration method that (i) performs joint affine and deformable coarse-to-fine registration within a single network, and (ii) embeds transformers into a NICE registration framework to model long-range relevance between images. Extensive experiments with seven public datasets show that our NICE-Trans outperforms state-of-the-art registration methods on both registration accuracy and runtime.

Mingyuan Meng, Michael Fulham, Dagan Feng et al.

Deformable image registration aims to find a dense non-linear spatial correspondence between a pair of images, which is a crucial step for many medical tasks such as tumor growth monitoring and population analysis. Recently, Deep Neural Networks (DNNs) have been widely recognized for their ability to perform fast end-to-end registration. However, DNN-based registration needs to explore the spatial information of each image and fuse this information to characterize spatial correspondence. This raises an essential question: what is the optimal fusion strategy to characterize spatial correspondence? Existing fusion strategies (e.g., early fusion, late fusion) were empirically designed to fuse information by manually defined prior knowledge, which inevitably constrains the registration performance within the limits of empirical designs. In this study, we depart from existing empirically-designed fusion strategies and develop a data-driven fusion strategy for deformable image registration. To achieve this, we propose an Automatic Fusion network (AutoFuse) that provides flexibility to fuse information at many potential locations within the network. A Fusion Gate (FG) module is also proposed to control how to fuse information at each potential network location based on training data. Our AutoFuse can automatically optimize its fusion strategy during training and can be generalizable to both unsupervised registration (without any labels) and semi-supervised registration (with weak labels provided for partial training data). Extensive experiments on two well-benchmarked medical registration tasks (inter- and intra-patient registration) with eight public datasets show that our AutoFuse outperforms state-of-the-art unsupervised and semi-supervised registration methods.

Mingyuan Meng, Lei Bi, Michael Fulham et al.

Survival prediction is crucial for cancer patients as it provides early prognostic information for treatment planning. Recently, deep survival models based on deep learning and medical images have shown promising performance for survival prediction. However, existing deep survival models are not well developed in utilizing multi-modality images (e.g., PET-CT) and in extracting region-specific information (e.g., the prognostic information in Primary Tumor (PT) and Metastatic Lymph Node (MLN) regions). In view of this, we propose a merging-diverging learning framework for survival prediction from multi-modality images. This framework has a merging encoder to fuse multi-modality information and a diverging decoder to extract region-specific information. In the merging encoder, we propose a Hybrid Parallel Cross-Attention (HPCA) block to effectively fuse multi-modality features via parallel convolutional layers and cross-attention transformers. In the diverging decoder, we propose a Region-specific Attention Gate (RAG) block to screen out the features related to lesion regions. Our framework is demonstrated on survival prediction from PET-CT images in Head and Neck (H&N) cancer, by designing an X-shape merging-diverging hybrid transformer network (named XSurv). Our XSurv combines the complementary information in PET and CT images and extracts the region-specific prognostic information in PT and MLN regions. Extensive experiments on the public dataset of HEad and neCK TumOR segmentation and outcome prediction challenge (HECKTOR 2022) demonstrate that our XSurv outperforms state-of-the-art survival prediction methods.

Yuxin Xue, Lei Bi, Yige Peng et al.

Positron emission tomography (PET) is a widely used, highly sensitive molecular imaging in clinical diagnosis. There is interest in reducing the radiation exposure from PET but also maintaining adequate image quality. Recent methods using convolutional neural networks (CNNs) to generate synthesized high-quality PET images from low-dose counterparts have been reported to be state-of-the-art for low-to-high image recovery methods. However, these methods are prone to exhibiting discrepancies in texture and structure between synthesized and real images. Furthermore, the distribution shift between low-dose PET and standard PET has not been fully investigated. To address these issues, we developed a self-supervised adaptive residual estimation generative adversarial network (SS-AEGAN). We introduce (1) An adaptive residual estimation mapping mechanism, AE-Net, designed to dynamically rectify the preliminary synthesized PET images by taking the residual map between the low-dose PET and synthesized output as the input, and (2) A self-supervised pre-training strategy to enhance the feature representation of the coarse generator. Our experiments with a public benchmark dataset of total-body PET images show that SS-AEGAN consistently outperformed the state-of-the-art synthesis methods with various dose reduction factors.

Lei Bi, Michael Fulham, Shaoli Song et al.

[18F]-Fluorodeoxyglucose (FDG) positron emission tomography - computed tomography (PET-CT) has become the imaging modality of choice for diagnosing many cancers. Co-learning complementary PET-CT imaging features is a fundamental requirement for automatic tumor segmentation and for developing computer aided cancer diagnosis systems. In this study, we propose a hyper-connected transformer (HCT) network that integrates a transformer network (TN) with a hyper connected fusion for multi-modality PET-CT images. The TN was leveraged for its ability to provide global dependencies in image feature learning, which was achieved by using image patch embeddings with a self-attention mechanism to capture image-wide contextual information. We extended the single-modality definition of TN with multiple TN based branches to separately extract image features. We also introduced a hyper connected fusion to fuse the contextual and complementary image features across multiple transformers in an iterative manner. Our results with two clinical datasets show that HCT achieved better performance in segmentation accuracy when compared to the existing methods.

Lei Bi, Michael Fulham, Jinman Kim

Automatic skin lesion segmentation methods based on fully convolutional networks (FCNs) are regarded as the state-of-the-art for accuracy. When there are, however, insufficient training data to cover all the variations in skin lesions, where lesions from different patients may have major differences in size/shape/texture, these methods failed to segment the lesions that have image characteristics, which are less common in the training datasets. FCN-based semi-automatic segmentation methods, which fuse user-inputs with high-level semantic image features derived from FCNs offer an ideal complement to overcome limitations of automatic segmentation methods. These semi-automatic methods rely on the automated state-of-the-art FCNs coupled with user-inputs for refinements, and therefore being able to tackle challenging skin lesions. However, there are a limited number of FCN-based semi-automatic segmentation methods and all these methods focused on early-fusion, where the first few convolutional layers are used to fuse image features and user-inputs and then derive fused image features for segmentation. For early-fusion based methods, because the user-input information can be lost after the first few convolutional layers, consequently, the user-input information will have limited guidance and constraint in segmenting the challenging skin lesions with inhomogeneous textures and fuzzy boundaries. Hence, in this work, we introduce a hyper-fusion network (HFN) to fuse the extracted user-inputs and image features over multiple stages. We separately extract complementary features which then allows for an iterative use of user-inputs along all the fusion stages to refine the segmentation. We evaluated our HFN on ISIC 2017, ISIC 2016 and PH2 datasets, and our results show that the HFN is more accurate and generalizable than the state-of-the-art methods.

Mingjian Li, Mingyuan Meng, Shuchang Ye et al.

Medical image segmentation is crucial in modern medical image analysis, which can aid into diagnosis of various disease conditions. Recently, language-guided segmentation methods have shown promising results in automating image segmentation where text reports are incorporated as guidance. These text reports, containing image impressions and insights given by clinicians, provides auxiliary guidance. However, these methods neglect the inherent pattern gaps between the two distinct modalities, which leads to sub-optimal image-text feature fusion without proper cross-modality feature alignments. Contrastive alignments are widely used to associate image-text semantics in representation learning; however, it has not been exploited to bridge the pattern gaps in language-guided segmentation that relies on subtle low level image details to represent diseases. Existing contrastive alignment methods typically algin high-level global image semantics without involving low-level, localized target information, and therefore fails to explore fine-grained text guidance for language-guided segmentation. In this study, we propose a language-guided segmentation network with Target-informed Multi-level Contrastive Alignments (TMCA). TMCA enables target-informed cross-modality alignments and fine-grained text guidance to bridge the pattern gaps in language-guided segmentation. Specifically, we introduce: 1) a target-sensitive semantic distance module that enables granular image-text alignment modelling, and 2) a multi-level alignment strategy that directs text guidance on low-level image features. In addition, a language-guided target enhancement module is proposed to leverage the aligned text to redirect attention to focus on critical localized image features. Extensive experiments on 4 image-text datasets, involving 3 medical imaging modalities, demonstrated that our TMCA achieved superior performances.

Andrew Zhang, Hao Wang, Shuchang Ye et al.

Patient motion during medical image acquisition causes blurring, ghosting, and distorts organs, which makes image interpretation challenging. Current state-of-the-art algorithms using Generative Adversarial Network (GAN)-based methods with their ability to learn the mappings between corrupted images and their ground truth via Structural Similarity Index Measure (SSIM) loss effectively generate motion-free images. However, we identified the following limitations: (i) they mainly focus on global structural characteristics and therefore overlook localized features that often carry critical pathological information, and (ii) the SSIM loss function struggles to handle images with varying pixel intensities, luminance factors, and variance. In this study, we propose Motion-Aware Image SYnthesis (MAISY) which initially characterize motion and then uses it for correction by: (a) leveraging the foundation model Segment Anything Model (SAM), to dynamically learn spatial patterns along anatomical boundaries where motion artifacts are most pronounced and, (b) introducing the Variance-Selective SSIM (VS-SSIM) loss which adaptively emphasizes spatial regions with high pixel variance to preserve essential anatomical details during artifact correction. Experiments on chest and head CT datasets demonstrate that our model outperformed the state-of-the-art counterparts, with Peak Signal-to-Noise Ratio (PSNR) increasing by 40%, SSIM by 10%, and Dice by 16%.

Mingyuan Meng, Michael Fulham, Lei Bi et al.

Deformable image registration is a fundamental requirement for medical image analysis. Recently, transformers have been widely used in deep learning-based registration methods for their ability to capture long-range dependency via self-attention (SA). However, the high computation and memory loads of SA (growing quadratically with the spatial resolution) hinder transformers from processing subtle textural information in high-resolution image features, e.g., at the full and half image resolutions. This limits deformable registration as the high-resolution textural information is crucial for finding precise pixel-wise correspondence between subtle anatomical structures. Cross-covariance Attention (XCA), as a "transposed" version of SA that operates across feature channels, has complexity growing linearly with the spatial resolution, providing the feasibility of capturing long-range dependency among high-resolution image features. However, existing XCA-based transformers merely capture coarse global long-range dependency, which are unsuitable for deformable image registration relying primarily on fine-grained local correspondence. In this study, we propose to improve existing deep learning-based registration methods by embedding a new XCA mechanism. To this end, we design an XCA-based transformer block optimized for deformable medical image registration, named Multi-Axis XCA (MAXCA). Our MAXCA serves as a general network block that can be embedded into various registration network architectures. It can capture both global and local long-range dependency among high-resolution image features by applying regional and dilated XCA in parallel via a multi-axis design. Extensive experiments on two well-benchmarked inter-/intra-patient registration tasks with seven public medical datasets demonstrate that our MAXCA block enables state-of-the-art registration performance.

Mingjian Li, Mingyuan Meng, Michael Fulham et al.

Medical image representations can be learned through medical vision-language contrastive learning (mVLCL) where medical imaging reports are used as weak supervision through image-text alignment. These learned image representations can be transferred to and benefit various downstream medical vision tasks such as disease classification and segmentation. Recent mVLCL methods attempt to align image sub-regions and the report keywords as local-matchings. However, these methods aggregate all local-matchings via simple pooling operations while ignoring the inherent relations between them. These methods therefore fail to reason between local-matchings that are semantically related, e.g., local-matchings that correspond to the disease word and the location word (semantic-relations), and also fail to differentiate such clinically important local-matchings from others that correspond to less meaningful words, e.g., conjunction words (importance-relations). Hence, we propose a mVLCL method that models the inter-matching relations between local-matchings via a relation-enhanced contrastive learning framework (RECLF). In RECLF, we introduce a semantic-relation reasoning module (SRM) and an importance-relation reasoning module (IRM) to enable more fine-grained report supervision for image representation learning. We evaluated our method using six public benchmark datasets on four downstream tasks, including segmentation, zero-shot classification, linear classification, and cross-modal retrieval. Our results demonstrated the superiority of our RECLF over the state-of-the-art mVLCL methods with consistent improvements across single-modal and cross-modal tasks. These results suggest that our RECLF, by modelling the inter-matching relations, can learn improved medical image representations with better generalization capabilities.

Mingyuan Meng, Bingxin Gu, Michael Fulham et al.

Survival prediction is a major concern for cancer management. Deep survival models based on deep learning have been widely adopted to perform end-to-end survival prediction from medical images. Recent deep survival models achieved promising performance by jointly performing tumor segmentation with survival prediction, where the models were guided to extract tumor-related information through Multi-Task Learning (MTL). However, these deep survival models have difficulties in exploring out-of-tumor prognostic information. In addition, existing deep survival models are unable to effectively leverage multi-modality images. Empirically-designed fusion strategies were commonly adopted to fuse multi-modality information via task-specific manually-designed networks, thus limiting the adaptability to different scenarios. In this study, we propose an Adaptive Multi-modality Segmentation-to-Survival model (AdaMSS) for survival prediction from PET/CT images. Instead of adopting MTL, we propose a novel Segmentation-to-Survival Learning (SSL) strategy, where our AdaMSS is trained for tumor segmentation and survival prediction sequentially in two stages. This strategy enables the AdaMSS to focus on tumor regions in the first stage and gradually expand its focus to include other prognosis-related regions in the second stage. We also propose a data-driven strategy to fuse multi-modality information, which realizes adaptive optimization of fusion strategies based on training data during training. With the SSL and data-driven fusion strategies, our AdaMSS is designed as an adaptive model that can self-adapt its focus regions and fusion strategy for different training stages. Extensive experiments with two large clinical datasets show that our AdaMSS outperforms state-of-the-art survival prediction methods.

Hoijoon Jung, Younhyun Jung, Michael Fulham et al.

Mixed reality head-mounted displays (mxdR-HMD) have the potential to visualize volumetric medical imaging data in holograms to provide a true sense of volumetric depth. An effective user interface, however, has yet to be thoroughly studied. Tangible user interfaces (TUIs) enable a tactile interaction with a hologram through an object. The object has physical properties indicating how it might be used with multiple degrees-of-freedom. We propose a TUI using a planar object (PO) for the holographic medical volume visualization and exploration. We refer to it as mxdR hologram slicer (mxdR-HS). Users can slice the hologram to examine particular regions of interest (ROIs) and intermix complementary data and annotations. The mxdR-HS introduces a novel real-time ad-hoc marker-less PO tracking method that works with any PO where corners are visible. The aim of mxdR-HS is to maintain minimum computational latency while preserving practical tracking accuracy to enable seamless TUI integration in the commercial mxdR-HMD, which has limited computational resources. We implemented the mxdR-HS on a commercial Microsoft HoloLens with a built-in depth camera. Our experimental results showed our mxdR-HS had a superior computational latency but marginally lower tracking accuracy than two marker-based tracking methods and resulted in enhanced computational latency and tracking accuracy than 10 marker-less tracking methods. Our mxdR-HS, in a medical environment, can be suggested as a visual guide to display complex volumetric medical imaging data.

Kai-Chieh Liang, Lei Bi, Ashnil Kumar et al.

Sequential whole-body 18F-Fluorodeoxyglucose (FDG) positron emission tomography (PET) scans are regarded as the imaging modality of choice for the assessment of treatment response in the lymphomas because they detect treatment response when there may not be changes on anatomical imaging. Any computerized analysis of lymphomas in whole-body PET requires automatic segmentation of the studies so that sites of disease can be quantitatively monitored over time. State-of-the-art PET image segmentation methods are based on convolutional neural networks (CNNs) given their ability to leverage annotated datasets to derive high-level features about the disease process. Such methods, however, focus on PET images from a single time-point and discard information from other scans or are targeted towards specific organs and cannot cater for the multiple structures in whole-body PET images. In this study, we propose a spatio-temporal 'dual-stream' neural network (ST-DSNN) to segment sequential whole-body PET scans. Our ST-DSNN learns and accumulates image features from the PET images done over time. The accumulated image features are used to enhance the organs / structures that are consistent over time to allow easier identification of sites of active lymphoma. Our results show that our method outperforms the state-of-the-art PET image segmentation methods.

Yige Peng, Lei Bi, Ashnil Kumar et al.

Distant metastases (DM) refer to the dissemination of tumors, usually, beyond the organ where the tumor originated. They are the leading cause of death in patients with soft-tissue sarcomas (STSs). Positron emission tomography-computed tomography (PET-CT) is regarded as the imaging modality of choice for the management of STSs. It is difficult to determine from imaging studies which STS patients will develop metastases. 'Radiomics' refers to the extraction and analysis of quantitative features from medical images and it has been employed to help identify such tumors. The state-of-the-art in radiomics is based on convolutional neural networks (CNNs). Most CNNs are designed for single-modality imaging data (CT or PET alone) and do not exploit the information embedded in PET-CT where there is a combination of an anatomical and functional imaging modality. Furthermore, most radiomic methods rely on manual input from imaging specialists for tumor delineation, definition and selection of radiomic features. This approach, however, may not be scalable to tumors with complex boundaries and where there are multiple other sites of disease. We outline a new 3D CNN to help predict DM in STS patients from PET-CT data. The 3D CNN uses a constrained feature learning module and a hierarchical multi-modality feature learning module that leverages the complementary information from the modalities to focus on semantically important regions. Our results on a public PET-CT dataset of STS patients show that multi-modal information improves the ability to identify those patients who develop DM. Further our method outperformed all other related state-of-the-art methods.

Xiaohang Fu, Lei Bi, Ashnil Kumar et al.

The identification of melanoma involves an integrated analysis of skin lesion images acquired using the clinical and dermoscopy modalities. Dermoscopic images provide a detailed view of the subsurface visual structures that supplement the macroscopic clinical images. Melanoma diagnosis is commonly based on the 7-point visual category checklist (7PC). The 7PC contains intrinsic relationships between categories that can aid classification, such as shared features, correlations, and the contributions of categories towards diagnosis. Manual classification is subjective and prone to intra- and interobserver variability. This presents an opportunity for automated methods to improve diagnosis. Current state-of-the-art methods focus on a single image modality and ignore information from the other, or do not fully leverage the complementary information from both modalities. Further, there is not a method to exploit the intercategory relationships in the 7PC. In this study, we address these issues by proposing a graph-based intercategory and intermodality network (GIIN) with two modules. A graph-based relational module (GRM) leverages intercategorical relations, intermodal relations, and prioritises the visual structure details from dermoscopy by encoding category representations in a graph network. The category embedding learning module (CELM) captures representations that are specialised for each category and support the GRM. We show that our modules are effective at enhancing classification performance using a public dataset of dermoscopy-clinical images, and show that our method outperforms the state-of-the-art at classifying the 7PC categories and diagnosis.

Mingyuan Meng, Lei Bi, Michael Fulham et al.

Deformable image registration is fundamental for many medical image analyses. A key obstacle for accurate image registration lies in image appearance variations such as the variations in texture, intensities, and noise. These variations are readily apparent in medical images, especially in brain images where registration is frequently used. Recently, deep learning-based registration methods (DLRs), using deep neural networks, have shown computational efficiency that is several orders of magnitude faster than traditional optimization-based registration methods (ORs). DLRs rely on a globally optimized network that is trained with a set of training samples to achieve faster registration. DLRs tend, however, to disregard the target-pair-specific optimization inherent in ORs and thus have degraded adaptability to variations in testing samples. This limitation is severe for registering medical images with large appearance variations, especially since few existing DLRs explicitly take into account appearance variations. In this study, we propose an Appearance Adjustment Network (AAN) to enhance the adaptability of DLRs to appearance variations. Our AAN, when integrated into a DLR, provides appearance transformations to reduce the appearance variations during registration. In addition, we propose an anatomy-constrained loss function through which our AAN generates anatomy-preserving transformations. Our AAN has been purposely designed to be readily inserted into a wide range of DLRs and can be trained cooperatively in an unsupervised and end-to-end manner. We evaluated our AAN with three state-of-the-art DLRs on three well-established public datasets of 3D brain magnetic resonance imaging (MRI). The results show that our AAN consistently improved existing DLRs and outperformed state-of-the-art ORs on registration accuracy, while adding a fractional computational load to existing DLRs.

Xiaohang Fu, Lei Bi, Ashnil Kumar et al.

Accurate characterisation of visual attributes such as spiculation, lobulation, and calcification of lung nodules is critical in cancer management. The characterisation of these attributes is often subjective, which may lead to high inter- and intra-observer variability. Furthermore, lung nodules are often heterogeneous in the cross-sectional image slices of a 3D volume. Current state-of-the-art methods that score multiple attributes rely on deep learning-based multi-task learning (MTL) schemes. These methods, however, extract shared visual features across attributes and then examine each attribute without explicitly leveraging their inherent intercorrelations. Furthermore, current methods either treat each slice with equal importance without considering their relevance or heterogeneity, which limits performance. In this study, we address these challenges with a new convolutional neural network (CNN)-based MTL model that incorporates multiple attention-based learning modules to simultaneously score 9 visual attributes of lung nodules in computed tomography (CT) image volumes. Our model processes entire nodule volumes of arbitrary depth and uses a slice attention module to filter out irrelevant slices. We also introduce cross-attribute and attribute specialisation attention modules that learn an optimal amalgamation of meaningful representations to leverage relationships between attributes. We demonstrate that our model outperforms previous state-of-the-art methods at scoring attributes using the well-known public LIDC-IDRI dataset of pulmonary nodules from over 1,000 patients. Our model also performs competitively when repurposed for benign-malignant classification. Our attention modules also provide easy-to-interpret weights that offer insights into the predictions of the model.

Xiaohang Fu, Lei Bi, Ashnil Kumar et al.

Multimodal positron emission tomography-computed tomography (PET-CT) is used routinely in the assessment of cancer. PET-CT combines the high sensitivity for tumor detection with PET and anatomical information from CT. Tumor segmentation is a critical element of PET-CT but at present, there is not an accurate automated segmentation method. Segmentation tends to be done manually by different imaging experts and it is labor-intensive and prone to errors and inconsistency. Previous automated segmentation methods largely focused on fusing information that is extracted separately from the PET and CT modalities, with the underlying assumption that each modality contains complementary information. However, these methods do not fully exploit the high PET tumor sensitivity that can guide the segmentation. We introduce a multimodal spatial attention module (MSAM) that automatically learns to emphasize regions (spatial areas) related to tumors and suppress normal regions with physiologic high-uptake. The resulting spatial attention maps are subsequently employed to target a convolutional neural network (CNN) for segmentation of areas with higher tumor likelihood. Our MSAM can be applied to common backbone architectures and trained end-to-end. Our experimental results on two clinical PET-CT datasets of non-small cell lung cancer (NSCLC) and soft tissue sarcoma (STS) validate the effectiveness of the MSAM in these different cancer types. We show that our MSAM, with a conventional U-Net backbone, surpasses the state-of-the-art lung tumor segmentation approach by a margin of 7.6% in Dice similarity coefficient (DSC).

Lei Bi, Jinman Kim, Tingwei Su et al.

The accurate classification of mass lesions in the adrenal glands (adrenal masses), detected with computed tomography (CT), is important for diagnosis and patient management. Adrenal masses can be benign or malignant and benign masses have varying prevalence. Classification methods based on convolutional neural networks (CNNs) are the state-of-the-art in maximizing inter-class differences in large medical imaging training datasets. The application of CNNs, to adrenal masses is challenging due to large intra-class variations, large inter-class similarities and imbalanced training data due to the size of the mass lesions. We developed a deep multi-scale resemblance network (DMRN) to overcome these limitations and leveraged paired CNNs to evaluate the intra-class similarities. We used multi-scale feature embedding to improve the inter-class separability by iteratively combining complementary information produced at different scales of the input to create structured feature descriptors. We augmented the training data with randomly sampled paired adrenal masses to reduce the influence of imbalanced training data. We used 229 CT scans of patients with adrenal masses for evaluation. In a five-fold cross-validation, our method had the best results (89.52% in accuracy) when compared to the state-of-the-art methods (p<0.05). We conducted a generalizability analysis of our method on the ImageCLEF 2016 competition dataset for medical subfigure classification, which consists of a training set of 6,776 images and a test set of 4,166 images across 30 classes. Our method achieved better classification performance (85.90% in accuracy) when compared to the existing methods and was competitive when compared with methods that require additional training data (1.47% lower in accuracy). Our DMRN sub-classified adrenal masses on CT and was superior to state-of-the-art approaches.

Yige Peng, Lei Bi, Michael Fulham et al.

'Radiomics' is a method that extracts mineable quantitative features from radiographic images. These features can then be used to determine prognosis, for example, predicting the development of distant metastases (DM). Existing radiomics methods, however, require complex manual effort including the design of hand-crafted radiomic features and their extraction and selection. Recent radiomics methods, based on convolutional neural networks (CNNs), also require manual input in network architecture design and hyper-parameter tuning. Radiomic complexity is further compounded when there are multiple imaging modalities, for example, combined positron emission tomography - computed tomography (PET-CT) where there is functional information from PET and complementary anatomical localization information from computed tomography (CT). Existing multi-modality radiomics methods manually fuse the data that are extracted separately. Reliance on manual fusion often results in sub-optimal fusion because they are dependent on an 'expert's' understanding of medical images. In this study, we propose a multi-modality neural architecture search method (MM-NAS) to automatically derive optimal multi-modality image features for radiomics and thus negate the dependence on a manual process. We evaluated our MM-NAS on the ability to predict DM using a public PET-CT dataset of patients with soft-tissue sarcomas (STSs). Our results show that our MM-NAS had a higher prediction accuracy when compared to state-of-the-art radiomics methods.

Ha Tran Hong Phan, Ashnil Kumar, David Feng et al.

Cell event detection in cell videos is essential for monitoring of cellular behavior over extended time periods. Deep learning methods have shown great success in the detection of cell events for their ability to capture more discriminative features of cellular processes compared to traditional methods. In particular, convolutional long short-term memory (LSTM) models, which exploits the changes in cell events observable in video sequences, is the state-of-the-art for mitosis detection in cell videos. However, their limitations are the determination of the input sequence length, which is often performed empirically, and the need for a large annotated training dataset which is expensive to prepare. We propose a novel semi-supervised method of optimal length detection for mitosis detection with two key contributions: (i) an unsupervised step for learning the spatial and temporal locations of cells in their normal stage and approximating the distribution of temporal lengths of cell events and, (ii) a step of inferring, from that distribution, an optimal input sequence length and a minimal number of annotated frames for training a LSTM model for each particular video. We evaluated our method in detecting mitosis in densely packed stem cells in a phase-contrast microscopy videos. Our experimental data prove that increasing the input sequence length of LSTM can lead to a decrease in performance. Our results also show that by approximating the optimal input sequence length of the tested video, a model trained with only 18 annotated frames achieved F1-scores of 0.880-0.907, which are 10% higher than those of other published methods with a full set of 110 training annotated frames.

Euijoon Ahn, Ashnil Kumar, Dagan Feng et al.

Medical image analysis using supervised deep learning methods remains problematic because of the reliance of deep learning methods on large amounts of labelled training data. Although medical imaging data repositories continue to expand there has not been a commensurate increase in the amount of annotated data. Hence, we propose a new unsupervised feature learning method that learns feature representations to then differentiate dissimilar medical images using an ensemble of different convolutional neural networks (CNNs) and K-means clustering. It jointly learns feature representations and clustering assignments in an end-to-end fashion. We tested our approach on a public medical dataset and show its accuracy was better than state-of-the-art unsupervised feature learning methods and comparable to state-of-the-art supervised CNNs. Our findings suggest that our method could be used to tackle the issue of the large volume of unlabelled data in medical imaging repositories.

Euijoon Ahn, Ashnil Kumar, Dagan Feng et al.

The accuracy and robustness of image classification with supervised deep learning are dependent on the availability of large-scale, annotated training data. However, there is a paucity of annotated data available due to the complexity of manual annotation. To overcome this problem, a popular approach is to use transferable knowledge across different domains by: 1) using a generic feature extractor that has been pre-trained on large-scale general images (i.e., transfer-learned) but which not suited to capture characteristics from medical images; or 2) fine-tuning generic knowledge with a relatively smaller number of annotated images. Our aim is to reduce the reliance on annotated training data by using a new hierarchical unsupervised feature extractor with a convolutional auto-encoder placed atop of a pre-trained convolutional neural network. Our approach constrains the rich and generic image features from the pre-trained domain to a sophisticated representation of the local image characteristics from the unannotated medical image domain. Our approach has a higher classification accuracy than transfer-learned approaches and is competitive with state-of-the-art supervised fine-tuned methods.

Lei Bi, Yuyu Guo, Qian Wang et al.

Automated segmentation of the optic cup and disk on retinal fundus images is fundamental for the automated detection / analysis of glaucoma. Traditional segmentation approaches depend heavily upon hand-crafted features and a priori knowledge of the user. As such, these methods are difficult to be adapt to the clinical environment. Recently, deep learning methods based on fully convolutional networks (FCNs) have been successful in resolving segmentation problems. However, the reliance on large annotated training data is problematic when dealing with medical images. If a sufficient amount of annotated training data to cover all possible variations is not available, FCNs do not provide accurate segmentation. In addition, FCNs have a large receptive field in the convolutional layers, and hence produce coarse outputs of boundaries. Hence, we propose a new fully automated method that we refer to as a dual-stage fully convolutional networks (DSFCN). Our approach leverages deep residual architectures and FCNs and learns and infers the location of the optic cup and disk in a step-wise manner with fine-grained details. During training, our approach learns from the training data and the estimated results derived from the previous iteration. The ability to learn from the previous iteration optimizes the learning of the optic cup and the disk boundaries. During testing (prediction), DSFCN uses test (input) images and the estimated probability map derived from previous iterations to gradually improve the segmentation accuracy. Our method achieved an average Dice co-efficient of 0.8488 and 0.9441 for optic cup and disk segmentation and an area under curve (AUC) of 0.9513 for glaucoma detection.

Ashnil Kumar, Michael Fulham, Dagan Feng et al.

The analysis of multi-modality positron emission tomography and computed tomography (PET-CT) images for computer aided diagnosis applications requires combining the sensitivity of PET to detect abnormal regions with anatomical localization from CT. Current methods for PET-CT image analysis either process the modalities separately or fuse information from each modality based on knowledge about the image analysis task. These methods generally do not consider the spatially varying visual characteristics that encode different information across the different modalities, which have different priorities at different locations. For example, a high abnormal PET uptake in the lungs is more meaningful for tumor detection than physiological PET uptake in the heart. Our aim is to improve fusion of the complementary information in multi-modality PET-CT with a new supervised convolutional neural network (CNN) that learns to fuse complementary information for multi-modality medical image analysis. Our CNN first encodes modality-specific features and then uses them to derive a spatially varying fusion map that quantifies the relative importance of each modality's features across different spatial locations. These fusion maps are then multiplied with the modality-specific feature maps to obtain a representation of the complementary multi-modality information at different locations, which can then be used for image analysis. We evaluated the ability of our CNN to detect and segment multiple regions with different fusion requirements using a dataset of PET-CT images of lung cancer. We compared our method to baseline techniques for multi-modality image fusion and segmentation. Our findings show that our CNN had a significantly higher foreground detection accuracy (99.29%, p < 0.05) than the fusion baselines and a significantly higher Dice score (63.85%) than recent PET-CT tumor segmentation methods.

Haozhe Jia, Yang Song, Donghao Zhang et al.

Advanced deep learning methods have been developed to conduct prostate MR volume segmentation in either a 2D or 3D fully convolutional manner. However, 2D methods tend to have limited segmentation performance, since large amounts of spatial information of prostate volumes are discarded during the slice-by-slice segmentation process; and 3D methods also have room for improvement, since they use isotropic kernels to perform 3D convolutions whereas most prostate MR volumes have anisotropic spatial resolution. Besides, the fully convolutional structural methods achieve good performance for localization issues but neglect the per-voxel classification for segmentation tasks. In this paper, we propose a 3D Global Convolutional Adversarial Network (3D GCA-Net) to address efficient prostate MR volume segmentation. We first design a 3D ResNet encoder to extract 3D features from prostate scans, and then develop the decoder, which is composed of a multi-scale 3D global convolutional block and a 3D boundary refinement block, to address the classification and localization issues simultaneously for volumetric segmentation. Additionally, we combine the encoder-decoder segmentation network with an adversarial network in the training phrase to enforce the contiguity of long-range spatial predictions. Throughout the proposed model, we use anisotropic convolutional processing for better feature learning on prostate MR scans. We evaluated our 3D GCA-Net model on two public prostate MR datasets and achieved state-of-the-art performances.

Euijoon Ahn, Jinman Kim, Ashnil Kumar et al.

The availability of large-scale annotated image datasets and recent advances in supervised deep learning methods enable the end-to-end derivation of representative image features that can impact a variety of image analysis problems. Such supervised approaches, however, are difficult to implement in the medical domain where large volumes of labelled data are difficult to obtain due to the complexity of manual annotation and inter- and intra-observer variability in label assignment. We propose a new convolutional sparse kernel network (CSKN), which is a hierarchical unsupervised feature learning framework that addresses the challenge of learning representative visual features in medical image analysis domains where there is a lack of annotated training data. Our framework has three contributions: (i) We extend kernel learning to identify and represent invariant features across image sub-patches in an unsupervised manner. (ii) We initialise our kernel learning with a layer-wise pre-training scheme that leverages the sparsity inherent in medical images to extract initial discriminative features. (iii) We adapt a multi-scale spatial pyramid pooling (SPP) framework to capture subtle geometric differences between learned visual features. We evaluated our framework in medical image retrieval and classification on three public datasets. Our results show that our CSKN had better accuracy when compared to other conventional unsupervised methods and comparable accuracy to methods that used state-of-the-art supervised convolutional neural networks (CNNs). Our findings indicate that our unsupervised CSKN provides an opportunity to leverage unannotated big data in medical imaging repositories.

Ha Tran Hong Phan, Ashnil Kumar, David Feng et al.

We propose an automatic unsupervised cell event detection and classification method, which expands convolutional Long Short-Term Memory (LSTM) neural networks, for cellular events in cell video sequences. Cells in images that are captured from various biomedical applications usually have different shapes and motility, which pose difficulties for the automated event detection in cell videos. Current methods to detect cellular events are based on supervised machine learning and rely on tedious manual annotation from investigators with specific expertise. So that our LSTM network could be trained in an unsupervised manner, we designed it with a branched structure where one branch learns the frequent, regular appearance and movements of objects and the second learns the stochastic events, which occur rarely and without warning in a cell video sequence. We tested our network on a publicly available dataset of densely packed stem cell phase-contrast microscopy images undergoing cell division. This dataset is considered to be more challenging that a dataset with sparse cells. We compared our method to several published supervised methods evaluated on the same dataset and to a supervised LSTM method with a similar design and configuration to our unsupervised method. We used an F1-score, which is a balanced measure for both precision and recall. Our results show that our unsupervised method has a higher or similar F1-score when compared to two fully supervised methods that are based on Hidden Conditional Random Fields (HCRF), and has comparable accuracy with the current best supervised HCRF-based method. Our method was generalizable as after being trained on one video it could be applied to videos where the cells were in different conditions. The accuracy of our unsupervised method approached that of its supervised counterpart.

Lei Bi, Jinman Kim, Ashnil Kumar et al.

Positron emission tomography (PET) image synthesis plays an important role, which can be used to boost the training data for computer aided diagnosis systems. However, existing image synthesis methods have problems in synthesizing the low resolution PET images. To address these limitations, we propose multi-channel generative adversarial networks (M-GAN) based PET image synthesis method. Different to the existing methods which rely on using low-level features, the proposed M-GAN is capable to represent the features in a high-level of semantic based on the adversarial learning concept. In addition, M-GAN enables to take the input from the annotation (label) to synthesize the high uptake regions e.g., tumors and from the computed tomography (CT) images to constrain the appearance consistency and output the synthetic PET images directly. Our results on 50 lung cancer PET-CT studies indicate that our method was much closer to the real PET images when compared with the existing methods.

Siqi Liu, Sidong Liu, Sonia Pujol et al.

We present a longitudinal MR simulation framework which simulates the future neurodegenerative progression by outputting the predicted follow-up MR image and the voxel-based morphometry (VBM) map. This framework expects the patients to have at least 2 historical MR images available. The longitudinal and cross-sectional VBM maps are extracted to measure the affinity between the target subject and the template subjects collected for simulation. Then the follow-up simulation is performed by resampling the latest available target MR image with a weighted sum of non-linear transformations derived from the best-matched templates. The leave-one-out strategy was used to compare different simulation methods. Compared to the state-of-the-art voxel-based method, our proposed morphometry-based simulation achieves better accuracy in most cases.