Assessing generalisability of deep learning-based polyp detection and segmentation methods through a computer vision challenge.

Polyps are well-known cancer precursors identified by colonoscopy. However, variability in their size, appearance, and location makes the detection of polyps challenging. Moreover, colonoscopy surveillance and removal of polyps are highly operator-dependent procedures and occur in a highly complex organ topology. There exists a high missed detection rate and incomplete removal of colonic polyps. To assist in clinical procedures and reduce missed rates, automated methods for detecting and segmenting polyps using machine learning have been achieved in past years. However, the major drawback in most of these methods is their ability to generalise to out-of-sample unseen datasets from different centres, populations, modalities, and acquisition systems. To test this hypothesis rigorously, we, together with expert gastroenterologists, curated a multi-centre and multi-population dataset acquired from six different colonoscopy systems and challenged the computational expert teams to develop robust automated detection and segmentation methods in a crowd-sourcing Endoscopic computer vision challenge. This work put forward rigorous generalisability tests and assesses the usability of devised deep learning methods in dynamic and actual clinical colonoscopy procedures. We analyse the results of four top performing teams for the detection task and five top performing teams for the segmentation task. Our analyses demonstrate that the top-ranking teams concentrated mainly on accuracy over the real-time performance required for clinical applicability. We further dissect the devised methods and provide an experiment-based hypothesis that reveals the need for improved generalisability to tackle diversity present in multi-centre datasets and routine clinical procedures.

CELLULAR, A Cell Autophagy Imaging Dataset.

Cells in living organisms are dynamic compartments that continuously respond to changes in their environment to maintain physiological homeostasis. While basal autophagy exists in cells to aid in the regular turnover of intracellular material, autophagy is also a critical cellular response to stress, such as nutritional depletion. Conversely, the deregulation of autophagy is linked to several diseases, such as cancer, and hence, autophagy constitutes a potential therapeutic target. Image analysis to follow autophagy in cells, especially on high-content screens, has proven to be a bottleneck. Machine learning (ML) algorithms have recently emerged as crucial in analyzing images to efficiently extract information, thus contributing to a better understanding of the questions at hand. This paper presents CELLULAR, an open dataset consisting of images of cells expressing the autophagy reporter mRFP-EGFP-Atg8a with cell-specific segmentation masks. Each cell is annotated into either basal autophagy, activated autophagy, or unknown. Furthermore, we introduce some preliminary experiments using the dataset that can be used as a baseline for future research.

Usefulness of Heat Map Explanations for Deep-Learning-Based Electrocardiogram Analysis.

Deep neural networks are complex machine learning models that have shown promising results in analyzing high-dimensional data such as those collected from medical examinations. Such models have the potential to provide fast and accurate medical diagnoses. However, the high complexity makes deep neural networks and their predictions difficult to understand. Providing model explanations can be a way of increasing the understanding of "black box" models and building trust. In this work, we applied transfer learning to develop a deep neural network to predict sex from electrocardiograms. Using the visual explanation method Grad-CAM, heat maps were generated from the model in order to understand how it makes predictions. To evaluate the usefulness of the heat maps and determine if the heat maps identified electrocardiogram features that could be recognized to discriminate sex, medical doctors provided feedback. Based on the feedback, we concluded that, in our setting, this mode of explainable artificial intelligence does not provide meaningful information to medical doctors and is not useful in the clinic. Our results indicate that improved explanation techniques that are tailored to medical data should be developed before deep neural networks can be applied in the clinic for diagnostic purposes.

VISEM-Tracking, a human spermatozoa tracking dataset.

A manual assessment of sperm motility requires microscopy observation, which is challenging due to the fast-moving spermatozoa in the field of view. To obtain correct results, manual evaluation requires extensive training. Therefore, computer-aided sperm analysis (CASA) has become increasingly used in clinics. Despite this, more data is needed to train supervised machine learning approaches in order to improve accuracy and reliability in the assessment of sperm motility and kinematics. In this regard, we provide a dataset called VISEM-Tracking with 20 video recordings of 30 seconds (comprising 29,196 frames) of wet semen preparations with manually annotated bounding-box coordinates and a set of sperm characteristics analyzed by experts in the domain. In addition to the annotated data, we provide unlabeled video clips for easy-to-use access and analysis of the data via methods such as self- or unsupervised learning. As part of this paper, we present baseline sperm detection performances using the YOLOv5 deep learning (DL) model trained on the VISEM-Tracking dataset. As a result, we show that the dataset can be used to train complex DL models to analyze spermatozoa.

GridHTM: Grid-Based Hierarchical Temporal Memory for Anomaly Detection in Videos.

The interest in video anomaly detection systems that can detect different types of anomalies, such as violent behaviours in surveillance videos, has gained traction in recent years. The current approaches employ deep learning to perform anomaly detection in videos, but this approach has multiple problems. For example, deep learning in general has issues with noise, concept drift, explainability, and training data volumes. Additionally, anomaly detection in itself is a complex task and faces challenges such as unknownness, heterogeneity, and class imbalance. Anomaly detection using deep learning is therefore mainly constrained to generative models such as generative adversarial networks and autoencoders due to their unsupervised nature; however, even they suffer from general deep learning issues and are hard to properly train. In this paper, we explore the capabilities of the Hierarchical Temporal Memory (HTM) algorithm to perform anomaly detection in videos, as it has favorable properties such as noise tolerance and online learning which combats concept drift. We introduce a novel version of HTM, named GridHTM, which is a grid-based HTM architecture specifically for anomaly detection in complex videos such as surveillance footage. We have tested GridHTM using the VIRAT video surveillance dataset, and the subsequent evaluation results and online learning capabilities prove the great potential of using our system for real-time unsupervised anomaly detection in complex videos.

SinGAN-Seg: Synthetic training data generation for medical image segmentation.

Analyzing medical data to find abnormalities is a time-consuming and costly task, particularly for rare abnormalities, requiring tremendous efforts from medical experts. Therefore, artificial intelligence has become a popular tool for the automatic processing of medical data, acting as a supportive tool for doctors. However, the machine learning models used to build these tools are highly dependent on the data used to train them. Large amounts of data can be difficult to obtain in medicine due to privacy reasons, expensive and time-consuming annotations, and a general lack of data samples for infrequent lesions. In this study, we present a novel synthetic data generation pipeline, called SinGAN-Seg, to produce synthetic medical images with corresponding masks using a single training image. Our method is different from the traditional generative adversarial networks (GANs) because our model needs only a single image and the corresponding ground truth to train. We also show that the synthetic data generation pipeline can be used to produce alternative artificial segmentation datasets with corresponding ground truth masks when real datasets are not allowed to share. The pipeline is evaluated using qualitative and quantitative comparisons between real data and synthetic data to show that the style transfer technique used in our pipeline significantly improves the quality of the generated data and our method is better than other state-of-the-art GANs to prepare synthetic images when the size of training datasets are limited. By training UNet++ using both real data and the synthetic data generated from the SinGAN-Seg pipeline, we show that the models trained on synthetic data have very close performances to those trained on real data when both datasets have a considerable amount of training data. In contrast, we show that synthetic data generated from the SinGAN-Seg pipeline improves the performance of segmentation models when training datasets do not have a considerable amount of data. All experiments were performed using an open dataset and the code is publicly available on GitHub.

On evaluation metrics for medical applications of artificial intelligence.

Clinicians and software developers need to understand how proposed machine learning (ML) models could improve patient care. No single metric captures all the desirable properties of a model, which is why several metrics are typically reported to summarize a model's performance. Unfortunately, these measures are not easily understandable by many clinicians. Moreover, comparison of models across studies in an objective manner is challenging, and no tool exists to compare models using the same performance metrics. This paper looks at previous ML studies done in gastroenterology, provides an explanation of what different metrics mean in the context of binary classification in the presented studies, and gives a thorough explanation of how different metrics should be interpreted. We also release an open source web-based tool that may be used to aid in calculating the most relevant metrics presented in this paper so that other researchers and clinicians may easily incorporate them into their research.

Meta-learning with implicit gradients in a few-shot setting for medical image segmentation.

Widely used traditional supervised deep learning methods require a large number of training samples but often fail to generalize on unseen datasets. Therefore, a more general application of any trained model is quite limited for medical imaging for clinical practice. Using separately trained models for each unique lesion category or a unique patient population will require sufficiently large curated datasets, which is not practical to use in a real-world clinical set-up. Few-shot learning approaches can not only minimize the need for an enormous number of reliable ground truth labels that are labour-intensive and expensive, but can also be used to model on a dataset coming from a new population. To this end, we propose to exploit an optimization-based implicit model agnostic meta-learning (iMAML) algorithm under few-shot settings for medical image segmentation. Our approach can leverage the learned weights from diverse but small training samples to perform analysis on unseen datasets with high accuracy. We show that, unlike classical few-shot learning approaches, our method improves generalization capability. To our knowledge, this is the first work that exploits iMAML for medical image segmentation and explores the strength of the model on scenarios such as meta-training on unique and mixed instances of lesion datasets. Our quantitative results on publicly available skin and polyp datasets show that the proposed method outperforms the naive supervised baseline model and two recent few-shot segmentation approaches by large margins. In addition, our iMAML approach shows an improvement of 2%-4% in dice score compared to its counterpart MAML for most experiments.

Impact of Image Resolution on Deep Learning Performance in Endoscopy Image Classification: An Experimental Study Using a Large Dataset of Endoscopic Images.

Recent trials have evaluated the efficacy of deep convolutional neural network (CNN)-based AI systems to improve lesion detection and characterization in endoscopy. Impressive results are achieved, but many medical studies use a very small image resolution to save computing resources at the cost of losing details. Today, no conventions between resolution and performance exist, and monitoring the performance of various CNN architectures as a function of image resolution provides insights into how subtleties of different lesions on endoscopy affect performance. This can help set standards for image or video characteristics for future CNN-based models in gastrointestinal (GI) endoscopy. This study examines the performance of CNNs on the HyperKvasir dataset, consisting of 10,662 images from 23 different findings. We evaluate two CNN models for endoscopic image classification under quality distortions with image resolutions ranging from 32 × 32 to 512 × 512 pixels. The performance is evaluated using two-fold cross-validation and F1-score, maximum Matthews correlation coefficient (MCC), precision, and sensitivity as metrics. Increased performance was observed with higher image resolution for all findings in the dataset. MCC was achieved at image resolutions between 512 × 512 pixels for classification for the entire dataset after including all subclasses. The highest performance was observed with an MCC value of 0.9002 when the models were trained on the highest resolution and tested on the same resolution. Different resolutions and their effect on CNNs are explored. We show that image resolution has a clear influence on the performance which calls for standards in the field in the future.

DeepFake electrocardiograms using generative adversarial networks are the beginning of the end for privacy issues in medicine.

Recent global developments underscore the prominent role big data have in modern medical science. But privacy issues constitute a prevalent problem for collecting and sharing data between researchers. However, synthetic data generated to represent real data carrying similar information and distribution may alleviate the privacy issue. In this study, we present generative adversarial networks (GANs) capable of generating realistic synthetic DeepFake 10-s 12-lead electrocardiograms (ECGs). We have developed and compared two methods, named WaveGAN* and Pulse2Pulse. We trained the GANs with 7,233 real normal ECGs to produce 121,977 DeepFake normal ECGs. By verifying the ECGs using a commercial ECG interpretation program (MUSE 12SL, GE Healthcare), we demonstrate that the Pulse2Pulse GAN was superior to the WaveGAN* to produce realistic ECGs. ECG intervals and amplitudes were similar between the DeepFake and real ECGs. Although these synthetic ECGs mimic the dataset used for creation, the ECGs are not linked to any individuals and may thus be used freely. The synthetic dataset will be available as open access for researchers at OSF.io and the DeepFake generator available at the Python Package Index (PyPI) for generating synthetic ECGs. In conclusion, we were able to generate realistic synthetic ECGs using generative adversarial neural networks on normal ECGs from two population studies, thereby addressing the relevant privacy issues in medical datasets.

Explaining deep neural networks for knowledge discovery in electrocardiogram analysis.

Deep learning-based tools may annotate and interpret medical data more quickly, consistently, and accurately than medical doctors. However, as medical doctors are ultimately responsible for clinical decision-making, any deep learning-based prediction should be accompanied by an explanation that a human can understand. We present an approach called electrocardiogram gradient class activation map (ECGradCAM), which is used to generate attention maps and explain the reasoning behind deep learning-based decision-making in ECG analysis. Attention maps may be used in the clinic to aid diagnosis, discover new medical knowledge, and identify novel features and characteristics of medical tests. In this paper, we showcase how ECGradCAM attention maps can unmask how a novel deep learning model measures both amplitudes and intervals in 12-lead electrocardiograms, and we show an example of how attention maps may be used to develop novel ECG features.

Kvasir-Capsule, a video capsule endoscopy dataset.

Artificial intelligence (AI) is predicted to have profound effects on the future of video capsule endoscopy (VCE) technology. The potential lies in improving anomaly detection while reducing manual labour. Existing work demonstrates the promising benefits of AI-based computer-assisted diagnosis systems for VCE. They also show great potential for improvements to achieve even better results. Also, medical data is often sparse and unavailable to the research community, and qualified medical personnel rarely have time for the tedious labelling work. We present Kvasir-Capsule, a large VCE dataset collected from examinations at a Norwegian Hospital. Kvasir-Capsule consists of 117 videos which can be used to extract a total of 4,741,504 image frames. We have labelled and medically verified 47,238 frames with a bounding box around findings from 14 different classes. In addition to these labelled images, there are 4,694,266 unlabelled frames included in the dataset. The Kvasir-Capsule dataset can play a valuable role in developing better algorithms in order to reach true potential of VCE technology.

A comprehensive analysis of classification methods in gastrointestinal endoscopy imaging.

Gastrointestinal (GI) endoscopy has been an active field of research motivated by the large number of highly lethal GI cancers. Early GI cancer precursors are often missed during the endoscopic surveillance. The high missed rate of such abnormalities during endoscopy is thus a critical bottleneck. Lack of attentiveness due to tiring procedures, and requirement of training are few contributing factors. An automatic GI disease classification system can help reduce such risks by flagging suspicious frames and lesions. GI endoscopy consists of several multi-organ surveillance, therefore, there is need to develop methods that can generalize to various endoscopic findings. In this realm, we present a comprehensive analysis of the Medico GI challenges: Medical Multimedia Task at MediaEval 2017, Medico Multimedia Task at MediaEval 2018, and BioMedia ACM MM Grand Challenge 2019. These challenges are initiative to set-up a benchmark for different computer vision methods applied to the multi-class endoscopic images and promote to build new approaches that could reliably be used in clinics. We report the performance of 21 participating teams over a period of three consecutive years and provide a detailed analysis of the methods used by the participants, highlighting the challenges and shortcomings of the current approaches and dissect their credibility for the use in clinical settings. Our analysis revealed that the participants achieved an improvement on maximum Mathew correlation coefficient (MCC) from 82.68% in 2017 to 93.98% in 2018 and 95.20% in 2019 challenges, and a significant increase in computational speed over consecutive years.

HyperKvasir, a comprehensive multi-class image and video dataset for gastrointestinal endoscopy.

Artificial intelligence is currently a hot topic in medicine. However, medical data is often sparse and hard to obtain due to legal restrictions and lack of medical personnel for the cumbersome and tedious process to manually label training data. These constraints make it difficult to develop systems for automatic analysis, like detecting disease or other lesions. In this respect, this article presents HyperKvasir, the largest image and video dataset of the gastrointestinal tract available today. The data is collected during real gastro- and colonoscopy examinations at Bærum Hospital in Norway and partly labeled by experienced gastrointestinal endoscopists. The dataset contains 110,079 images and 374 videos, and represents anatomical landmarks as well as pathological and normal findings. The total number of images and video frames together is around 1 million. Initial experiments demonstrate the potential benefits of artificial intelligence-based computer-assisted diagnosis systems. The HyperKvasir dataset can play a valuable role in developing better algorithms and computer-assisted examination systems not only for gastro- and colonoscopy, but also for other fields in medicine.

Machine Learning-Based Analysis of Sperm Videos and Participant Data for Male Fertility Prediction.

Methods for automatic analysis of clinical data are usually targeted towards a specific modality and do not make use of all relevant data available. In the field of male human reproduction, clinical and biological data are not used to its fullest potential. Manual evaluation of a semen sample using a microscope is time-consuming and requires extensive training. Furthermore, the validity of manual semen analysis has been questioned due to limited reproducibility, and often high inter-personnel variation. The existing computer-aided sperm analyzer systems are not recommended for routine clinical use due to methodological challenges caused by the consistency of the semen sample. Thus, there is a need for an improved methodology. We use modern and classical machine learning techniques together with a dataset consisting of 85 videos of human semen samples and related participant data to automatically predict sperm motility. Used techniques include simple linear regression and more sophisticated methods using convolutional neural networks. Our results indicate that sperm motility prediction based on deep learning using sperm motility videos is rapid to perform and consistent. Adding participant data did not improve the algorithms performance. In conclusion, machine learning-based automatic analysis may become a valuable tool in male infertility investigation and research.