Will Transformers change gastrointestinal endoscopic image analysis? A comparative analysis between CNNs and Transformers, in terms of performance, robustness and generalization.

Gastrointestinal endoscopic image analysis presents significant challenges, such as considerable variations in quality due to the challenging in-body imaging environment, the often-subtle nature of abnormalities with low interobserver agreement, and the need for real-time processing. These challenges pose strong requirements on the performance, generalization, robustness and complexity of deep learning-based techniques in such safety-critical applications. While Convolutional Neural Networks (CNNs) have been the go-to architecture for endoscopic image analysis, recent successes of the Transformer architecture in computer vision raise the possibility to update this conclusion. To this end, we evaluate and compare clinically relevant performance, generalization and robustness of state-of-the-art CNNs and Transformers for neoplasia detection in Barrett's esophagus. We have trained and validated several top-performing CNNs and Transformers on a total of 10,208 images (2,079 patients), and tested on a total of 7,118 images (998 patients) across multiple test sets, including a high-quality test set, two internal and two external generalization test sets, and a robustness test set. Furthermore, to expand the scope of the study, we have conducted the performance and robustness comparisons for colonic polyp segmentation (Kvasir-SEG) and angiodysplasia detection (Giana). The results obtained for featured models across a wide range of training set sizes demonstrate that Transformers achieve comparable performance as CNNs on various applications, show comparable or slightly improved generalization capabilities and offer equally strong resilience and robustness against common image corruptions and perturbations. These findings confirm the viability of the Transformer architecture, particularly suited to the dynamic nature of endoscopic video analysis, characterized by fluctuating image quality, appearance and equipment configurations in transition from hospital to hospital. The code is made publicly available at: https://github.com/BONS-AI-VCA-AMC/Endoscopy-CNNs-vs-Transformers.

Foundation models in gastrointestinal endoscopic AI: Impact of architecture, pre-training approach and data efficiency.

Pre-training deep learning models with large data sets of natural images, such as ImageNet, has become the standard for endoscopic image analysis. This approach is generally superior to training from scratch, due to the scarcity of high-quality medical imagery and labels. However, it is still unknown whether the learned features on natural imagery provide an optimal starting point for the downstream medical endoscopic imaging tasks. Intuitively, pre-training with imagery closer to the target domain could lead to better-suited feature representations. This study evaluates whether leveraging in-domain pre-training in gastrointestinal endoscopic image analysis has potential benefits compared to pre-training on natural images. To this end, we present a dataset comprising of 5,014,174 gastrointestinal endoscopic images from eight different medical centers (GastroNet-5M), and exploit self-supervised learning with SimCLRv2, MoCov2 and DINO to learn relevant features for in-domain downstream tasks. The learned features are compared to features learned on natural images derived with multiple methods, and variable amounts of data and/or labels (e.g. Billion-scale semi-weakly supervised learning and supervised learning on ImageNet-21k). The effects of the evaluation is performed on five downstream data sets, particularly designed for a variety of gastrointestinal tasks, for example, GIANA for angiodyplsia detection and Kvasir-SEG for polyp segmentation. The findings indicate that self-supervised domain-specific pre-training, specifically using the DINO framework, results into better performing models compared to any supervised pre-training on natural images. On the ResNet50 and Vision-Transformer-small architectures, utilizing self-supervised in-domain pre-training with DINO leads to an average performance boost of 1.63% and 4.62%, respectively, on the downstream datasets. This improvement is measured against the best performance achieved through pre-training on natural images within any of the evaluated frameworks. Moreover, the in-domain pre-trained models also exhibit increased robustness against distortion perturbations (noise, contrast, blur, etc.), where the in-domain pre-trained ResNet50 and Vision-Transformer-small with DINO achieved on average 1.28% and 3.55% higher on the performance metrics, compared to the best performance found for pre-trained models on natural images. Overall, this study highlights the importance of in-domain pre-training for improving the generic nature, scalability and performance of deep learning for medical image analysis. The GastroNet-5M pre-trained weights are made publicly available in our repository: huggingface.co/tgwboers/GastroNet-5M_Pretrained_Weights.

Investigating and Improving Latent Density Segmentation Models for Aleatoric Uncertainty Quantification in Medical Imaging.

Data uncertainties, such as sensor noise, occlusions or limitations in the acquisition method can introduce irreducible ambiguities in images, which result in varying, yet plausible, semantic hypotheses. In Machine Learning, this ambiguity is commonly referred to as aleatoric uncertainty. In image segmentation, latent density models can be utilized to address this problem. The most popular approach is the Probabilistic U-Net (PU-Net), which uses latent Normal densities to optimize the conditional data log-likelihood Evidence Lower Bound. In this work, we demonstrate that the PU-Net latent space is severely sparse and heavily under-utilized. To address this, we introduce mutual information maximization and entropy-regularized Sinkhorn Divergence in the latent space to promote homogeneity across all latent dimensions, effectively improving gradient-descent updates and latent space informativeness. Our results show that by applying this on public datasets of various clinical segmentation problems, our proposed methodology receives up to 11% performance gains compared against preceding latent variable models for probabilistic segmentation on the Hungarian-Matched Intersection over Union. The results indicate that encouraging a homogeneous latent space significantly improves latent density modeling for medical image segmentation.

Defining tumor growth in vestibular schwannomas: a volumetric inter-observer variability study in contrast-enhanced T1-weighted MRI.

PURPOSE: For patients with vestibular schwannomas (VS), a conservative observational approach is increasingly used. Therefore, the need for accurate and reliable volumetric tumor monitoring is important. Currently, a volumetric cutoff of 20% increase in tumor volume is widely used to define tumor growth in VS. The study investigates the tumor volume dependency on the limits of agreement (LoA) for volumetric measurements of VS by means of an inter-observer study. METHODS: This retrospective study included 100 VS patients who underwent contrast-enhanced T1-weighted MRI. Five observers volumetrically annotated the images. Observer agreement and reliability was measured using the LoA, estimated using the limits of agreement with the mean (LOAM) method, and the intraclass correlation coefficient (ICC). RESULTS: The 100 patients had a median average tumor volume of 903 mm3 (IQR: 193-3101). Patients were divided into four volumetric size categories based on tumor volume quartile. The smallest tumor volume quartile showed a LOAM relative to the mean of 26.8% (95% CI: 23.7-33.6), whereas for the largest tumor volume quartile this figure was found to be 7.3% (95% CI: 6.5-9.7) and when excluding peritumoral cysts: 4.8% (95% CI: 4.2-6.2). CONCLUSION: Agreement limits within volumetric annotation of VS are affected by tumor volume, since the LoA improves with increasing tumor volume. As a result, for tumors larger than 200 mm3, growth can reliably be detected at an earlier stage, compared to the currently widely used cutoff of 20%. However, for very small tumors, growth should be assessed with higher agreement limits than previously thought.

Appropriate trust in artificial intelligence for the optical diagnosis of colorectal polyps: The role of human/artificial intelligence interaction.

BACKGROUND AND AIMS: Computer-aided diagnosis (CADx) for the optical diagnosis of colorectal polyps is thoroughly investigated. However, studies on human-artificial intelligence interaction are lacking. Our aim was to investigate endoscopists' trust in CADx by evaluating whether communicating a calibrated algorithm confidence score improved trust. METHODS: Endoscopists optically diagnosed 60 colorectal polyps. Initially, endoscopists diagnosed the polyps without CADx assistance (initial diagnosis). Immediately afterward, the same polyp was again shown with a CADx prediction: either only a prediction (benign or premalignant) or a prediction accompanied by a calibrated confidence score (0-100). A confidence score of 0 indicated a benign prediction, 100 a (pre)malignant prediction. In half of the polyps, CADx was mandatory, and for the other half, CADx was optional. After reviewing the CADx prediction, endoscopists made a final diagnosis. Histopathology was used as the gold standard. Endoscopists' trust in CADx was measured as CADx prediction utilization: the willingness to follow CADx predictions when the endoscopists initially disagreed with the CADx prediction. RESULTS: Twenty-three endoscopists participated. Presenting CADx predictions increased the endoscopists' diagnostic accuracy (69.3% initial vs 76.6% final diagnosis, P < .001). The CADx prediction was used in 36.5% (n = 183 of 501) disagreements. Adding a confidence score led to lower CADx prediction utilization, except when the confidence score surpassed 60. Mandatory CADx decreased CADx prediction utilization compared to optional CADx. Appropriate trust-using correct or disregarding incorrect CADx predictions-was 48.7% (n = 244 of 501). CONCLUSIONS: Appropriate trust was common, and CADx prediction utilization was highest for the optional CADx without confidence scores. These results express the importance of a better understanding of human-artificial intelligence interaction.

Estimating Surgical Urethral Length on Intraoperative Robot-Assisted Prostatectomy Images Using Artificial Intelligence Anatomy Recognition.

Objective: To construct a convolutional neural network (CNN) model that can recognize and delineate anatomic structures on intraoperative video frames of robot-assisted radical prostatectomy (RARP) and to use these annotations to predict the surgical urethral length (SUL). Background: Urethral dissection during RARP impacts patient urinary incontinence (UI) outcomes, and requires extensive training. Large differences exist between incontinence outcomes of different urologists and hospitals. Also, surgeon experience and education are critical toward optimal outcomes. Therefore, new approaches are warranted. SUL is associated with UI. Artificial intelligence (AI) surgical image segmentation using a CNN could automate SUL estimation and contribute toward future AI-assisted RARP and surgeon guidance. Methods: Eighty-eight intraoperative RARP videos between June 2009 and September 2014 were collected from a single center. Two hundred sixty-four frames were annotated according to prostate, urethra, ligated plexus, and catheter. Thirty annotated images from different RARP videos were used as a test data set. The dice (similarity) coefficient (DSC) and 95th percentile Hausdorff distance (Hd95) were used to determine model performance. SUL was calculated using the catheter as a reference. Results: The DSC of the best performing model were 0.735 and 0.755 for the catheter and urethra classes, respectively, with a Hd95 of 29.27 and 72.62, respectively. The model performed moderately on the ligated plexus and prostate. The predicted SUL showed a mean difference of 0.64 to 1.86 mm difference vs human annotators, but with significant deviation (standard deviation = 3.28-3.56). Conclusion: This study shows that an AI image segmentation model can predict vital structures during RARP urethral dissection with moderate to fair accuracy. SUL estimation derived from it showed large deviations and outliers when compared with human annotators, but with a small mean difference (<2 mm). This is a promising development for further research on AI-assisted RARP.

Robustness evaluation of deep neural networks for endoscopic image analysis: Insights and strategies.

Computer-aided detection and diagnosis systems (CADe/CADx) in endoscopy are commonly trained using high-quality imagery, which is not representative for the heterogeneous input typically encountered in clinical practice. In endoscopy, the image quality heavily relies on both the skills and experience of the endoscopist and the specifications of the system used for screening. Factors such as poor illumination, motion blur, and specific post-processing settings can significantly alter the quality and general appearance of these images. This so-called domain gap between the data used for developing the system and the data it encounters after deployment, and the impact it has on the performance of deep neural networks (DNNs) supportive endoscopic CAD systems remains largely unexplored. As many of such systems, for e.g. polyp detection, are already being rolled out in clinical practice, this poses severe patient risks in particularly community hospitals, where both the imaging equipment and experience are subject to considerable variation. Therefore, this study aims to evaluate the impact of this domain gap on the clinical performance of CADe/CADx for various endoscopic applications. For this, we leverage two publicly available data sets (KVASIR-SEG and GIANA) and two in-house data sets. We investigate the performance of commonly-used DNN architectures under synthetic, clinically calibrated image degradations and on a prospectively collected dataset including 342 endoscopic images of lower subjective quality. Additionally, we assess the influence of DNN architecture and complexity, data augmentation, and pretraining techniques for improved robustness. The results reveal a considerable decline in performance of 11.6% (±1.5) as compared to the reference, within the clinically calibrated boundaries of image degradations. Nevertheless, employing more advanced DNN architectures and self-supervised in-domain pre-training effectively mitigate this drop to 7.7% (±2.03). Additionally, these enhancements yield the highest performance on the manually collected test set including images with lower subjective quality. By comprehensively assessing the robustness of popular DNN architectures and training strategies across multiple datasets, this study provides valuable insights into their performance and limitations for endoscopic applications. The findings highlight the importance of including robustness evaluation when developing DNNs for endoscopy applications and propose strategies to mitigate performance loss.

Computer-aided diagnosis improves characterization of Barrett's neoplasia by general endoscopists (with video).

BACKGROUND AND AIMS: Characterization of visible abnormalities in patients with Barrett's esophagus (BE) can be challenging, especially for inexperienced endoscopists. This results in suboptimal diagnostic accuracy and poor interobserver agreement. Computer-aided diagnosis (CADx) systems may assist endoscopists. We aimed to develop, validate, and benchmark a CADx system for BE neoplasia. METHODS: The CADx system received pretraining with ImageNet and then consecutive domain-specific pretraining with GastroNet, which includes 5 million endoscopic images. It was subsequently trained and internally validated using 1758 narrow-band imaging (NBI) images of early BE neoplasia (352 patients) and 1838 NBI images of nondysplastic BE (173 patients) from 8 international centers. CADx was tested prospectively on corresponding image and video test sets with 30 cases (20 patients) of BE neoplasia and 60 cases (31 patients) of nondysplastic BE. The test set was benchmarked by 44 general endoscopists in 2 phases (phase 1, no CADx assistance; phase 2, with CADx assistance). Ten international BE experts provided additional benchmark performance. RESULTS: Stand-alone sensitivity and specificity of the CADx system were 100% and 98% for images and 93% and 96% for videos, respectively. CADx outperformed general endoscopists without CADx assistance in terms of sensitivity (P = .04). Sensitivity and specificity of general endoscopists increased from 84% to 96% and 90% to 98% with CAD assistance (P < .001). CADx assistance increased endoscopists' confidence in characterization (P < .001). CADx performance was similar to that of the BE experts. CONCLUSIONS: CADx assistance significantly increased characterization performance of BE neoplasia by general endoscopists to the level of expert endoscopists. The use of this CADx system may thereby improve daily Barrett surveillance.

The use of a real-time computer-aided detection system for visible lesions in the Barrett's esophagus during live endoscopic procedures: a pilot study (with video).

BACKGROUND AND AIMS: This pilot study evaluated the performance of a recently developed computer-aided detection (CADe) system for Barrett's neoplasia during live endoscopic procedures. METHODS: Fifteen patients with a visible lesion and 15 without were included in this study. A CAD-assisted workflow was used that included a slow pullback video recording of the entire Barrett's segment with live CADe assistance, followed by CADe-assisted level-based video recordings every 2 cm of the Barrett's segment. Outcomes were per-patient and per-level diagnostic accuracy of the CAD-assisted workflow, in which the primary outcome was per-patient in vivo CADe sensitivity. RESULTS: In the per-patient analyses, the CADe system detected all visible lesions (sensitivity 100%). Per-patient CADe specificity was 53%. Per-level sensitivity and specificity of the CADe-assisted workflow were 100% and 73%, respectively. CONCLUSIONS: In this pilot study, detection by the CADe system of all potentially neoplastic lesions in Barrett's esophagus was comparable to that of an expert endoscopist. Continued refinement of the system may improve specificity. External validation in larger multicenter studies is planned. (Clinical trial registration number: NCT05628441.).

Advancing 6-DoF Instrument Pose Estimation in Variable X-Ray Imaging Geometries.

Accurate 6-DoF pose estimation of surgical instruments during minimally invasive surgeries can substantially improve treatment strategies and eventual surgical outcome. Existing deep learning methods have achieved accurate results, but they require custom approaches for each object and laborious setup and training environments often stretching to extensive simulations, whilst lacking real-time computation. We propose a general-purpose approach of data acquisition for 6-DoF pose estimation tasks in X-ray systems, a novel and general purpose YOLOv5-6D pose architecture for accurate and fast object pose estimation and a complete method for surgical screw pose estimation under acquisition geometry consideration from a monocular cone-beam X-ray image. The proposed YOLOv5-6D pose model achieves competitive results on public benchmarks whilst being considerably faster at 42 FPS on GPU. In addition, the method generalizes across varying X-ray acquisition geometry and semantic image complexity to enable accurate pose estimation over different domains. Finally, the proposed approach is tested for bone-screw pose estimation for computer-aided guidance during spine surgeries. The model achieves a 92.41% by the 0.1·d ADD-S metric, demonstrating a promising approach for enhancing surgical precision and patient outcomes. The code for YOLOv5-6D is publicly available at https://github.com/cviviers/YOLOv5-6D-Pose.

Computed Tomography-Based Radiomics Using Tumor and Vessel Features to Assess Resectability in Cancer of the Pancreatic Head.

The preoperative prediction of resectability pancreatic ductal adenocarcinoma (PDAC) is challenging. This retrospective single-center study examined tumor and vessel radiomics to predict the resectability of PDAC in chemo-naïve patients. The tumor and adjacent arteries and veins were segmented in the portal-venous phase of contrast-enhanced CT scans, and radiomic features were extracted. Features were selected via stability and collinearity testing, and least absolute shrinkage and selection operator application (LASSO). Three models, using tumor features, vessel features, and a combination of both, were trained with the training set (N = 86) to predict resectability. The results were validated with the test set (N = 15) and compared to the multidisciplinary team's (MDT) performance. The vessel-features-only model performed best, with an AUC of 0.92 and sensitivity and specificity of 97% and 73%, respectively. Test set validation showed a sensitivity and specificity of 100% and 88%, respectively. The combined model was as good as the vessel model (AUC = 0.91), whereas the tumor model showed poor performance (AUC = 0.76). The MDT's prediction reached a sensitivity and specificity of 97% and 84% for the training set and 88% and 100% for the test set, respectively. Our clinician-independent vessel-based radiomics model can aid in predicting resectability and shows performance comparable to that of the MDT. With these encouraging results, improved, automated, and generalizable models can be developed that reduce workload and can be applied in non-expert hospitals.

Towards a robust and compact deep learning system for primary detection of early Barrett's neoplasia: Initial image-based results of training on a multi-center retrospectively collected data set.

INTRODUCTION: Endoscopic detection of early neoplasia in Barrett's esophagus is difficult. Computer Aided Detection (CADe) systems may assist in neoplasia detection. The aim of this study was to report the first steps in the development of a CADe system for Barrett's neoplasia and to evaluate its performance when compared with endoscopists. METHODS: This CADe system was developed by a consortium, consisting of the Amsterdam University Medical Center, Eindhoven University of Technology, and 15 international hospitals. After pretraining, the system was trained and validated using 1.713 neoplastic (564 patients) and 2.707 non-dysplastic Barrett's esophagus (NDBE; 665 patients) images. Neoplastic lesions were delineated by 14 experts. The performance of the CADe system was tested on three independent test sets. Test set 1 (50 neoplastic and 150 NDBE images) contained subtle neoplastic lesions representing challenging cases and was benchmarked by 52 general endoscopists. Test set 2 (50 neoplastic and 50 NDBE images) contained a heterogeneous case-mix of neoplastic lesions, representing distribution in clinical practice. Test set 3 (50 neoplastic and 150 NDBE images) contained prospectively collected imagery. The main outcome was correct classification of the images in terms of sensitivity. RESULTS: The sensitivity of the CADe system on test set 1 was 84%. For general endoscopists, sensitivity was 63%, corresponding to a neoplasia miss-rate of one-third of neoplastic lesions and a potential relative increase in neoplasia detection of 33% for CADe-assisted detection. The sensitivity of the CADe system on test sets 2 and 3 was 100% and 88%, respectively. The specificity of the CADe system varied for the three test sets between 64% and 66%. CONCLUSION: This study describes the first steps towards the establishment of an unprecedented data infrastructure for using machine learning to improve the endoscopic detection of Barrett's neoplasia. The CADe system detected neoplasia reliably and outperformed a large group of endoscopists in terms of sensitivity.

Optical Biopsy of Dysplasia in Barrett's Oesophagus Assisted by Artificial Intelligence.

Optical biopsy in Barrett's oesophagus (BE) using endocytoscopy (EC) could optimize endoscopic screening. However, the identification of dysplasia is challenging due to the complex interpretation of the highly detailed images. Therefore, we assessed whether using artificial intelligence (AI) as second assessor could help gastroenterologists in interpreting endocytoscopic BE images. First, we prospectively videotaped 52 BE patients with EC. Then we trained and tested the AI pm distinct datasets drawn from 83,277 frames, developed an endocytoscopic BE classification system, and designed online training and testing modules. We invited two successive cohorts for these online modules: 10 endoscopists to validate the classification system and 12 gastroenterologists to evaluate AI as second assessor by providing six of them with the option to request AI assistance. Training the endoscopists in the classification system established an improved sensitivity of 90.0% (+32.67%, p < 0.001) and an accuracy of 77.67% (+13.0%, p = 0.020) compared with the baseline. However, these values deteriorated at follow-up (-16.67%, p < 0.001 and -8.0%, p = 0.009). Contrastingly, AI-assisted gastroenterologists maintained high sensitivity and accuracy at follow-up, subsequently outperforming the unassisted gastroenterologists (+20.0%, p = 0.025 and +12.22%, p = 0.05). Thus, best diagnostic scores for the identification of dysplasia emerged through human-machine collaboration between trained gastroenterologists with AI as the second assessor. Therefore, AI could support clinical implementation of optical biopsies through EC.

Development of a CT-Compatible, Anthropomorphic Skull and Brain Phantom for Neurosurgical Planning, Training, and Simulation.

BACKGROUND: Neurosurgical procedures are complex and require years of training and experience. Traditional training on human cadavers is expensive, requires facilities and planning, and raises ethical concerns. Therefore, the use of anthropomorphic phantoms could be an excellent substitute. The aim of the study was to design and develop a patient-specific 3D-skull and brain model with realistic CT-attenuation suitable for conventional and augmented reality (AR)-navigated neurosurgical simulations. METHODS: The radiodensity of materials considered for the skull and brain phantoms were investigated using cone beam CT (CBCT) and compared to the radiodensities of the human skull and brain. The mechanical properties of the materials considered were tested in the laboratory and subsequently evaluated by clinically active neurosurgeons. Optimization of the phantom for the intended purposes was performed in a feedback cycle of tests and improvements. RESULTS: The skull, including a complete representation of the nasal cavity and skull base, was 3D printed using polylactic acid with calcium carbonate. The brain was cast using a mixture of water and coolant, with 4 wt% polyvinyl alcohol and 0.1 wt% barium sulfate, in a mold obtained from segmentation of CBCT and T1 weighted MR images from a cadaver. The experiments revealed that the radiodensities of the skull and brain phantoms were 547 and 38 Hounsfield units (HU), as compared to real skull bone and brain tissues with values of around 1300 and 30 HU, respectively. As for the mechanical properties testing, the brain phantom exhibited a similar elasticity to real brain tissue. The phantom was subsequently evaluated by neurosurgeons in simulations of endonasal skull-base surgery, brain biopsies, and external ventricular drain (EVD) placement and found to fulfill the requirements of a surgical phantom. CONCLUSIONS: A realistic and CT-compatible anthropomorphic head phantom was designed and successfully used for simulated augmented reality-led neurosurgical procedures. The anatomic details of the skull base and brain were realistically reproduced. This phantom can easily be manufactured and used for surgical training at a low cost.

Noise Reduction in CT Using Learned Wavelet-Frame Shrinkage Networks.

Encoding-decoding (ED) CNNs have demonstrated state-of-the-art performance for noise reduction over the past years. This has triggered the pursuit of better understanding the inner workings of such architectures, which has led to the theory of deep convolutional framelets (TDCF), revealing important links between signal processing and CNNs. Specifically, the TDCF demonstrates that ReLU CNNs induce low-rankness, since these models often do not satisfy the necessary redundancy to achieve perfect reconstruction (PR). In contrast, this paper explores CNNs that do meet the PR conditions. We demonstrate that in these type of CNNs soft shrinkage and PR can be assumed. Furthermore, based on our explorations we propose the learned wavelet-frame shrinkage network, or LWFSN and its residual counterpart, the rLWFSN. The ED path of the (r)LWFSN complies with the PR conditions, while the shrinkage stage is based on the linear expansion of thresholds proposed Blu and Luisier. In addition, the LWFSN has only a fraction of the training parameters (<1%) of conventional CNNs, very small inference times, low memory footprint, while still achieving performance close to state-of-the-art alternatives, such as the tight frame (TF) U-Net and FBPConvNet, in low-dose CT denoising.

Double-Camera Fusion System for Animal-Position Awareness in Farming Pens.

In livestock breeding, continuous and objective monitoring of animals is manually unfeasible due to the large scale of breeding and expensive labour. Computer vision technology can generate accurate and real-time individual animal or animal group information from video surveillance. However, the frequent occlusion between animals and changes in appearance features caused by varying lighting conditions makes single-camera systems less attractive. We propose a double-camera system and image registration algorithms to spatially fuse the information from different viewpoints to solve these issues. This paper presents a deformable learning-based registration framework, where the input image pairs are initially linearly pre-registered. Then, an unsupervised convolutional neural network is employed to fit the mapping from one view to another, using a large number of unlabelled samples for training. The learned parameters are then used in a semi-supervised network and fine-tuned with a small number of manually annotated landmarks. The actual pixel displacement error is introduced as a complement to an image similarity measure. The performance of the proposed fine-tuned method is evaluated on real farming datasets and demonstrates significant improvement in lowering the registration errors than commonly used feature-based and intensity-based methods. This approach also reduces the registration time of an unseen image pair to less than 0.5 s. The proposed method provides a high-quality reference processing step for improving subsequent tasks such as multi-object tracking and behaviour recognition of animals for further analysis.

Medical Instrument Segmentation in 3D US by Hybrid Constrained Semi-Supervised Learning.

Medical instrument segmentation in 3D ultrasound is essential for image-guided intervention. However, to train a successful deep neural network for instrument segmentation, a large number of labeled images are required, which is expensive and time-consuming to obtain. In this article, we propose a semi-supervised learning (SSL) framework for instrument segmentation in 3D US, which requires much less annotation effort than the existing methods. To achieve the SSL learning, a Dual-UNet is proposed to segment the instrument. The Dual-UNet leverages unlabeled data using a novel hybrid loss function, consisting of uncertainty and contextual constraints. Specifically, the uncertainty constraints leverage the uncertainty estimation of the predictions of the UNet, and therefore improve the unlabeled information for SSL training. In addition, contextual constraints exploit the contextual information of the training images, which are used as the complementary information for voxel-wise uncertainty estimation. Extensive experiments on multiple ex-vivo and in-vivo datasets show that our proposed method achieves Dice score of about 68.6%-69.1% and the inference time of about 1 sec. per volume. These results are better than the state-of-the-art SSL methods and the inference time is comparable to the supervised approaches.

Local and Distributed Machine Learning for Inter-hospital Data Utilization: An Application for TAVI Outcome Prediction.

Background: Machine learning models have been developed for numerous medical prognostic purposes. These models are commonly developed using data from single centers or regional registries. Including data from multiple centers improves robustness and accuracy of prognostic models. However, data sharing between multiple centers is complex, mainly because of regulations and patient privacy issues. Objective: We aim to overcome data sharing impediments by using distributed ML and local learning followed by model integration. We applied these techniques to develop 1-year TAVI mortality estimation models with data from two centers without sharing any data. Methods: A distributed ML technique and local learning followed by model integration was used to develop models to predict 1-year mortality after TAVI. We included two populations with 1,160 (Center A) and 631 (Center B) patients. Five traditional ML algorithms were implemented. The results were compared to models created individually on each center. Results: The combined learning techniques outperformed the mono-center models. For center A, the combined local XGBoost achieved an AUC of 0.67 (compared to a mono-center AUC of 0.65) and, for center B, a distributed neural network achieved an AUC of 0.68 (compared to a mono-center AUC of 0.64). Conclusion: This study shows that distributed ML and combined local models techniques, can overcome data sharing limitations and result in more accurate models for TAVI mortality estimation. We have shown improved prognostic accuracy for both centers and can also be used as an alternative to overcome the problem of limited amounts of data when creating prognostic models.

Conditional Generative Adversarial Networks for low-dose CT image denoising aiming at preservation of critical image content.

X-ray Computed Tomography (CT) is an imaging modality where patients are exposed to potentially harmful ionizing radiation. To limit patient risk, reduced-dose protocols are desirable, which inherently lead to an increased noise level in the reconstructed CT scans. Consequently, noise reduction algorithms are indispensable in the reconstruction processing chain. In this paper, we propose to leverage a conditional Generative Adversarial Networks (cGAN) model, to translate CT images from low-to-routine dose. However, when aiming to produce realistic images, such generative models may alter critical image content. Therefore, we propose to employ a frequency-based separation of the input prior to applying the cGAN model, in order to limit the cGAN to high-frequency bands, while leaving low-frequency bands untouched. The results of the proposed method are compared to a state-of-the-art model within the cGAN model as well as in a single-network setting. The proposed method generates visually superior results compared to the single-network model and the cGAN model in terms of quality of texture and preservation of fine structural details. It also appeared that the PSNR, SSIM and TV metrics are less important than a careful visual evaluation of the results. The obtained results demonstrate the relevance of defining and separating the input image into desired and undesired content, rather than blindly denoising entire images. This study shows promising results for further investigation of generative models towards finding a reliable deep learning-based noise reduction algorithm for low-dose CT acquisition.

Automatic image and text-based description for colorectal polyps using BASIC classification.

Colorectal polyps (CRP) are precursor lesions of colorectal cancer (CRC). Correct identification of CRPs during in-vivo colonoscopy is supported by the endoscopist's expertise and medical classification models. A recent developed classification model is the Blue light imaging Adenoma Serrated International Classification (BASIC) which describes the differences between non-neoplastic and neoplastic lesions acquired with blue light imaging (BLI). Computer-aided detection (CADe) and diagnosis (CADx) systems are efficient at visually assisting with medical decisions but fall short at translating decisions into relevant clinical information. The communication between machine and medical expert is of crucial importance to improve diagnosis of CRP during in-vivo procedures. In this work, the combination of a polyp image classification model and a language model is proposed to develop a CADx system that automatically generates text comparable to the human language employed by endoscopists. The developed system generates equivalent sentences as the human-reference and describes CRP images acquired with white light (WL), blue light imaging (BLI) and linked color imaging (LCI). An image feature encoder and a BERT module are employed to build the AI model and an external test set is used to evaluate the results and compute the linguistic metrics. The experimental results show the construction of complete sentences with an established metric scores of BLEU-1 = 0.67, ROUGE-L = 0.83 and METEOR = 0.50. The developed CADx system for automatic CRP image captioning facilitates future advances towards automatic reporting and may help reduce time-consuming histology assessment.