ABSTRACT
Scarcity of labels for medical images is a significant barrier for training representation learning approaches based on deep neural networks. This limitation is also present when using imaging data collected during routine clinical care stored in picture archiving communication systems (PACS), as these data rarely have attached the high-quality labels required for medical image computing tasks. However, medical images extracted from PACS are commonly coupled with descriptive radiology reports that contain significant information and could be leveraged to pre-train imaging models, which could serve as starting points for further task-specific fine-tuning. In this work, we perform a head-to-head comparison of three different self-supervised strategies to pre-train the same imaging model on 3D brain computed tomography angiogram (CTA) images, with large vessel occlusion (LVO) detection as the downstream task. These strategies evaluate two natural language processing (NLP) approaches, one to extract 100 explicit radiology concepts (Rad-SpatialNet) and the other to create general-purpose radiology reports embeddings (DistilBERT). In addition, we experiment with learning radiology concepts directly or by using a recent self-supervised learning approach (CLIP) that learns by ranking the distance between language and image vector embeddings. The LVO detection task was selected because it requires 3D imaging data, is clinically important, and requires the algorithm to learn outputs not explicitly stated in the radiology report. Pre-training was performed on an unlabeled dataset containing 1,542 3D CTA - reports pairs. The downstream task was tested on a labeled dataset of 402 subjects for LVO. We find that the pre-training performed with CLIP-based strategies improve the performance of the imaging model to detect LVO compared to a model trained only on the labeled data. The best performance was achieved by pre-training using the explicit radiology concepts and CLIP strategy.
ABSTRACT
Idiopathic Rem sleep Behavior Disorder (iRBD) is a significant biomarker for the development of alpha-synucleinopathies, such as Parkinson's disease (PD) or Dementia with Lewy bodies (DLB). Methods to identify patterns in iRBD patients can help in the prediction of the future conversion to these diseases during the long prodromal phase when symptoms are non-specific. These methods are essential for disease management and clinical trial recruitment. Brain PET scans with 18F-FDG PET radiotracers have recently shown promise, however, the scarcity of longitudinal data and PD/DLB conversion information makes the use of representation learning approaches such as deep convolutional networks not feasible if trained in a supervised manner. In this work, we propose a self-supervised learning strategy to learn features by comparing the brain hemispheres of iRBD non-convertor subjects, which allows for pre-training a convolutional network on a small data regimen. We introduce a loss function called hemisphere dissimilarity loss (HDL), which extends the Barlow Twins loss, that promotes the creation of invariant and non-redundant features for brain hemispheres of the same subject, and the opposite for hemispheres of different subjects. This loss enables the pre-training of a network without any information about the disease, which is then used to generate full brain feature vectors that are fine-tuned to two downstream tasks: follow-up conversion, and the type of conversion (PD or DLB) using baseline 18F-FDG PET. In our results, we find that the HDL outperforms the variational autoencoder with different forms of inputs.
ABSTRACT
The foveal avascular zone (FAZ) is a retinal area devoid of capillaries and associated with multiple retinal pathologies and visual acuity. Optical Coherence Tomography Angiography (OCT-A) is a very effective means of visualizing retinal vascular and avascular areas, but its use remains limited to research settings due to its complex optics limiting availability. On the other hand, fundus photography is widely available and often adopted in population studies. In this work, we test the feasibility of estimating the FAZ from fundus photos using three different approaches. The first two approaches rely on pixel-level and image-level FAZ information to segment FAZ pixels and regress FAZ area, respectively. The third is a training mask-free pipeline combining saliency maps with an active contours approach to segment FAZ pixels while being trained on image-level measures of the FAZ areas. This enables training FAZ segmentation methods without manual alignment of fundus and OCT-A images, a time-consuming process, which limits the dataset that can be used for training. Segmentation methods trained on pixel-level labels and image-level labels had good agreement with masks from a human grader (respectively DICE of 0.45 and 0.4). Results indicate the feasibility of using fundus images as a proxy to estimate the FAZ when angiography data is not available.
ABSTRACT
A correction has been published and is appended to both the HTML and PDF versions of this paper. The error has not been fixed in the paper.
ABSTRACT
Parkinson's disease (PD) is a slowly progressing neurodegenerative disease with early manifestation of motor signs. Objective measurements of motor signs are of vital importance for diagnosing, monitoring and developing disease modifying therapies, particularly for the early stages of the disease when putative neuroprotective treatments could stop neurodegeneration. Current medical practice has limited tools to routinely monitor PD motor signs with enough frequency and without undue burden for patients and the healthcare system. In this paper, we present data indicating that the routine interaction with computer keyboards can be used to detect motor signs in the early stages of PD. We explore a solution that measures the key hold times (the time required to press and release a key) during the normal use of a computer without any change in hardware and converts it to a PD motor index. This is achieved by the automatic discovery of patterns in the time series of key hold times using an ensemble regression algorithm. This new approach discriminated early PD groups from controls with an AUC = 0.81 (n = 42/43; mean age = 59.0/60.1; women = 43%/60%;PD/controls). The performance was comparable or better than two other quantitative motor performance tests used clinically: alternating finger tapping (AUC = 0.75) and single key tapping (AUC = 0.61).
Subject(s)
Models, Biological , Motor Activity , Parkinson Disease/diagnosis , Parkinson Disease/physiopathology , User-Computer Interface , HumansABSTRACT
Modern digital devices and appliances are capable of monitoring the timing of button presses, or finger interactions in general, with a sub-millisecond accuracy. However, the massive amount of high resolution temporal information that these devices could collect is currently being discarded. Multiple studies have shown that the act of pressing a button triggers well defined brain areas which are known to be affected by motor-compromised conditions. In this study, we demonstrate that the daily interaction with a computer keyboard can be employed as means to observe and potentially quantify psychomotor impairment. We induced a psychomotor impairment via a sleep inertia paradigm in 14 healthy subjects, which is detected by our classifier with an Area Under the ROC Curve (AUC) of 0.93/0.91. The detection relies on novel features derived from key-hold times acquired on standard computer keyboards during an uncontrolled typing task. These features correlate with the progression to psychomotor impairment (p < 0.001) regardless of the content and language of the text typed, and perform consistently with different keyboards. The ability to acquire longitudinal measurements of subtle motor changes from a digital device without altering its functionality may allow for early screening and follow-up of motor-compromised neurodegenerative conditions, psychological disorders or intoxication at a negligible cost in the general population.
Subject(s)
Fingers/physiology , Psychomotor Disorders/diagnosis , Adult , Algorithms , Area Under Curve , Computers , Female , Healthy Volunteers , Humans , Male , ROC Curve , Young AdultABSTRACT
Geographic Atrophy (GA) of the retinal pigment epithelium (RPE) is an advanced form of atrophic age-related macular degeneration (AMD) and is responsible for about 20% of AMD-related legal blindness in the United States. Two different imaging modalities for retinas, infrared imaging and autofluorescence imaging, serve as interesting complimentary technologies for highlighting GA. In this work we explore the use of neural network classifiers in performing segmentation of GA in registered infrared (IR) and autofluorescence (AF) images. Our segmentation achieved a performance level of 82.5% sensitivity and 92.9% specificity on a per-pixel basis using hold-one-out validation testing. The algorithm, feature extraction, data set and experimental results are discussed and shown.
Subject(s)
Geographic Atrophy/pathology , Learning , Neural Networks, Computer , Retina/pathology , HumansABSTRACT
The creation of an automatic diabetic retinopathy screening system using retina cameras is currently receiving considerable interest in the medical imaging community. The detection of microaneurysms is a key element in this effort. In this work, we propose a new microaneurysms segmentation technique based on a novel application of the radon transform, which is able to identify these lesions without any previous knowledge of the retina morphological features and with minimal image preprocessing. The algorithm has been evaluated on the Retinopathy Online Challenge public dataset, and its performance compares with the best current techniques. The performance is particularly good at low false positive ratios, which makes it an ideal candidate for diabetic retinopathy screening systems.
Subject(s)
Algorithms , Aneurysm/pathology , Diabetic Retinopathy/pathology , Image Interpretation, Computer-Assisted/methods , Pattern Recognition, Automated/methods , Retinal Artery/pathology , Retinoscopy/methods , Humans , Image Enhancement/methods , Reproducibility of Results , Sensitivity and SpecificityABSTRACT
The automated detection of diabetic retinopathy and other eye diseases in images of the retina has great promise as a low-cost method for broad-based screening. Many systems in the literature which perform automated detection include a quality estimation step and physiological feature detection, including the vascular tree and the optic nerve / macula location. In this work, we study the robustness of an automated disease detection method with respect to the accuracy of the optic nerve location and the quality of the images obtained as judged by a quality estimation algorithm. The detection algorithm features microaneurysm and exudate detection followed by feature extraction on the detected population to describe the overall retina image. Labeled images of retinas ground-truthed to disease states are used to train a supervised learning algorithm to identify the disease state of the retina image and exam set. Under the restrictions of high confidence optic nerve detections and good quality imagery, the system achieves a sensitivity and specificity of 94.8% and 78.7% with area-under-curve of 95.3%. Analysis of the effect of constraining quality and the distinction between mild non-proliferative diabetic retinopathy, normal retina images, and more severe disease states is included.
Subject(s)
Algorithms , Diabetic Retinopathy/pathology , Image Enhancement/methods , Image Interpretation, Computer-Assisted/methods , Pattern Recognition, Automated/methods , Retinoscopy/methods , Humans , Reproducibility of Results , Sensitivity and SpecificityABSTRACT
Age related Macular Degeneration (AMD) is a disease of the retina associated with aging. AMD progression in patients is characterized by drusen, pigmentation changes, and geographic atrophy, which can be seen using fundus imagery. The level of AMD is characterized by standard scaling methods, which can be somewhat subjective in practice. In this work we propose a statistical image processing approach to segment drusen with the ultimate goal of characterizing the AMD progression in a data set of longitudinal images. The method characterizes retinal structures with a statistical model of the colors in the retina image. When comparing the segmentation results of the method between longitudinal images with known AMD progression and those without, the method detects progression in our longitudinal data set with an area under the receiver operating characteristics curve of 0.99.
Subject(s)
Macular Degeneration/diagnosis , Macular Degeneration/pathology , Retinal Drusen/diagnosis , Retinal Drusen/pathology , Algorithms , Atrophy/pathology , Colorimetry/methods , Databases, Factual , Disease Progression , Fundus Oculi , Humans , Image Processing, Computer-Assisted , Models, Statistical , Neural Networks, Computer , Normal Distribution , Pigmentation , ROC Curve , Retina/pathologyABSTRACT
The projected increase in diabetes in the United States and worldwide has created a need for broad-based, inexpensive screening for diabetic retinopathy (DR), an eye disease which can lead to vision impairment. A telemedicine network with retina cameras and automated quality control, physiological feature location, and lesion / anomaly detection is a low-cost way of achieving broad-based screening. In this work we report on the effect of quality estimation on an optic nerve (ON) detection method with a confidence metric. We report on an improvement of the method using a data set from an ophthalmologist practice then show the results of the method as a function of image quality on a set of images from an on-line telemedicine network collected in Spring 2009 and another broad-based screening program. We show that the fusion method, combined with quality estimation processing, can improve detection performance and also provide a method for utilizing a physician-in-the-loop for images that may exceed the capabilities of automated processing.
Subject(s)
Diabetic Retinopathy/pathology , Image Interpretation, Computer-Assisted/methods , Optic Nerve/pathology , Radiology Information Systems/organization & administration , Retinoscopy/methods , Telemedicine/methods , Humans , Reproducibility of Results , Sensitivity and SpecificityABSTRACT
A great effort of the research community is geared towards the creation of an automatic screening system able to promptly detect diabetic retinopathy with the use of fundus cameras. In addition, there are some documented approaches for automatically judging the image quality. We propose a new set of features independent of field of view or resolution to describe the morphology of the patient's vessels. Our initial results suggest that these features can be used to estimate the image quality in a time one order of magnitude shorter than previous techniques.