Machine learning can accurately detect abnormal aortic valves in CMR

Background

Improving the efficiency of CMR by acquiring fewer, and more targeted sequences, would improve the diagnostic yield and reduce patient recalls. An AI-assisted clinical decision support system (CDSS) could deliver this efficiency using adaptive scanning protocols which replicate the expertise of highly trained clinicians. Normal aortic valve anatomy on the three-chamber (3CH) cine CMR is a guide to rationalising subsequent sequences, and therefore is a suitable base case for developing an AI-CDSS for CMR.

Purpose

We propose a machine learning approach to differentiate between normal and abnormal aortic valves from the 3CH cine.

Methods

We curated a unique expert-annotated dataset of 1221 frames from eighty CMR studies. For each frame, AV landmarks (two hinge points and two leaflets), and stenotic and regurgitant jets were labelled by three cardiologists.

We then tested two AI models (Figure 1) to detect these AV abnormalities: A) a convolutional neural network (CNN), and B) a random forest approach.

A) Using heat map regression, the AV was localised, and the jets (if present) were identified as pathological curves. We then tracked and quantified the curves in the estimated heatmaps based on their proximity, the length, orientation and angle with respect to the hinge points.

B) We used a random forest approach to classify cases as normal or abnormal by using the characteristics of estimated pathological curves obtained from the heat map regression output.

We trained and evaluated our models on an unseen dataset of 1017 CMR studies obtained from different scanner types across three NHS hospitals. Each CMR study report was manually assigned a binary ground truth label for a normal or abnormal AV. In total 496/1017 patients had an abnormal AV. Of those abnormal cases, 184 patients had aortic stenosis, 222 aortic regurgitation and 90 cases had mixed valve disease.

We assessed the classification performance of our method with accuracy and an F1 score – a composite of precision and recall, where 1 is perfect; and heatmap regression performance for curves with mean absolute error.

Results

This machine learning approach classified abnormal aortic valves with good agreement to the ground truth labels with mean accuracy of 0.93 (representing approximately 451/496 patients) and mean F1 score of 0.91. The AV hinge points were localised with a mean distance error of 3.5 pixels. This was despite the small size of expert labelled data.

Conclusion

This machine learning solution successfully differentiated between normal and abnormal aortic valves from routine 3CH cine CMR views. More labelled datasets will enable further classification of pathology and severity, and greater accuracy. Our results represent an important stepping stone towards an AI-assisted CDSS for CMR.

Funding Acknowledgement

Type of funding sources: Other. Main funding source(s): This work was supported by the UKRI CDT in AI for Healthcare http://ai4health.io

Framework for large-scale automatic curation of heterogeneous cardiac MRI (ACUR MRI)

Funding Acknowledgements

Type of funding sources: Public grant(s) – National budget only. Main funding source(s): UKRI CDT in AI for Healthcare http://ai4health.io and British Heart Foundation

Background 

Data curation is an important process that structures and organises data, supporting research and the development of artificial intelligence models. However, manually curating a large volume of medical data is a time-consuming, repetitive and costly process that puts additional strain on clinical experts. The curation becomes more complex and demanding as more data sources are used. This leads to an introduction of disparity in the data structure and protocols. 

Purpose 

Here, we propose an automatic framework to curate large volumes of heterogenous cardiac MRI scans acquired across different sites and scanner vendors. Our framework requires minimal expert involvement throughout and works directly on DICOM images from the scanner or PACS.  The resulting structured standardised data allow for straightforward image analysis, hypothesis testing and the training and application of artificial intelligence models. 

Methods 

It is broken down into three main components

anonymisation, cataloguing and outlier detection (see Figure 1). Anonymisation automatically removes any identifiable patient information from the DICOM image attributes. These data are replaced with anonymised labels, whilst maintaining relevant longitudinal information from each patient. DICOM attributes are also used to automatically group the different images according to imaging sequence (e.g. CINE, Delayed-Enhancement, T1 maps), acquisition geometry (e.g. short-axis, 2-chamber, 4-chamber) and imaging attributes (e.g. slice thickness, TE, TR), for easier querying. The sorting characteristics are flexible and can easily be defined by the user. Finally, we detect and flag, for subsequent manual inspection, any outliers within those groups, based on the similarity levels of chosen DICOM attributes. This framework additionally offers interactive image visualisation to allow users to assess its performance in real time. 

Results 

We tested the performance of ACUR CMRI on 26,668 CMR image series (723,531 images) from 858 patient examinations, which took place across two sites in four different scanners. With an average execution time per patient of 100 seconds, ACUR was able to sort imaging data with 1191 different sequence names into 43 categories. The framework can be freely downloaded from https://bitbucket.org/cmr-ai-working-group/acur/. 

Conclusions 

We present ACUR, an automatic framework to curate large volumes of heterogeneous cardiac MRI data. We show how it can quickly and automatically curate data, grouping it according to desired imaging characteristics defined in DICOM attributes. The proposed framework is flexible and ideally suited as a pre-processing tool for large biomedical imaging data studies.

Automatic estimation of left atrial function from short axis CINE-MRI using machine learning

Introduction

The importance of atrial mechanical dysfunction in atrial and ventricular pathologies is becoming increasingly recognised. Although machine learning (ML) tools have the ability to automatically estimate atrial function, to date ML techniques have not been used to automatically estimate atrial volumes and functional parameters directly from short axis CINE MRI.

Purpose

We introduce a convolutional neural network (CNN) to automatically segment the left atria (LA) in CINE-MRI. As a demonstration of the clinical utility of this technique, we calculated LA and left ventricular (LV) ejection fractions automatically from CINE images.

Methods

Short axis CINE MRI stacks, covering both ventricles and atria, were obtained in a 1.5T Philips Ingenia scanner. A 2D bSSFP ECG-gated protocol was used (FA=60°, TE/TR=1.5/2.9 ms), typical FOV =385 x 310 x 150 mm3, acquisition matrix = 172 x 140, slice thickness = 10 mm, reconstructed with resolution 1.25 x 1.25 x 10 mm3, 30–50 cardiac phases. Images were collected from 37 AF patients in sinus rythm at the time of scan (31–72 years old, 75% male, 18 with paroxysmal AF (PAF), 19 with persistent AF (persAF)).

To automatically segment the LA, we used a dedicated CNN that follows a U-Net architecture and was trained in 715 images of the LA, manually segmented by an expert. Data augmentation techniques that included noise addition and linear and non-linear image transforms were also used to increase the training dataset. Ventricular structures, including the LV blood pool, were automatically segmented in these images using a CNN previously trained for this task.

Volumetric time plots of LA and LV volume were produced and used to automatically compute maximal and minimal volumes, from which LA and LV ejection fractions (EFs) were assessed. A Bland-Altman analysis compared these automatically computed LA volumes and LA EFs with clinical manual estimates from the same scanning session.

Results

The CNN achieved very good quality LA segmentations when compared to manual ones (Fig a,b): Dice coefficients (0.90±0.07), median contour distances (0.50±1.12mm) and Hausdorff distances (6.70±6.16mm). Bland-Altman analyses show very good agreement between automatic and manual LA volumes and EFs (Fig e). A moderate linear correlation between LA and LV EFs in AF patients was found (Fig d). The measured LA EF was higher for PAF (29±8%) than PersAF patients (21±11%), although non-significantly (t-test p-value: 0.10).

Conclusions

We present a reliable automatic method to perform LA segmentations from CINE MRI across the entire cardiac cycle. This approachs opens up the possibility of automatically calculating more sophisticated biomarkers of LA function which take into account information about LA volumes across the entire cardiac cycle, including biomarkers of LA booster pump function.

Figure 1

Funding Acknowledgement

Type of funding source: Foundation. Main funding source(s): British Heart Foundation; EPSRC/Wellcome Centre for Medical Engineering

Correction to: Current concept in upright spinal MRI

Unfortunately, the legend of Fig. 5 was incorrectly published in original publication. The corrected legend is given below.

HMM-based lexicon-driven and lexicon-free word recognition for online handwritten Indic scripts

Research for recognizing online handwritten words in Indic scripts is at its early stages when compared to Latin and Oriental scripts. In this paper, we address this problem specifically for two major Indic scripts--Devanagari and Tamil. In contrast to previous approaches, the techniques we propose are largely data driven and script independent. We propose two different techniques for word recognition based on Hidden Markov Models (HMM): lexicon driven and lexicon free. The lexicon-driven technique models each word in the lexicon as a sequence of symbol HMMs according to a standard symbol writing order derived from the phonetic representation. The lexicon-free technique uses a novel Bag-of-Symbols representation of the handwritten word that is independent of symbol order and allows rapid pruning of the lexicon. On handwritten Devanagari word samples featuring both standard and nonstandard symbol writing orders, a combination of lexicon-driven and lexicon-free recognizers significantly outperforms either of them used in isolation. In contrast, most Tamil word samples feature the standard symbol order, and the lexicon-driven recognizer outperforms the lexicon free one as well as their combination. The best recognition accuracies obtained for 20,000 word lexicons are 87.13 percent for Devanagari when the two recognizers are combined, and 91.8 percent for Tamil using the lexicon-driven technique.

Computer algorithms for the automated measurement of retinal arteriolar diameters

AIMS

Quantification of retinal vascular change is difficult and manual measurements of vascular features are slow and subject to observer bias. These problems may be overcome using computer algorithms. Three automated methods and a manual method for measurement of arteriolar diameters from digitised red-free retinal photographs were compared.

METHODS

60 diameters (in pixels) measured by manual identification of vessel edges in red-free retinal images were compared with diameters measured by (1) fitting vessel intensity profiles to a double Gaussian function by non-linear regression, (2) a standard edge detection algorithm (Sobel), and (3) determination of points of maximum intensity variation by a sliding linear regression filter (SLRF). Method agreement was analysed using Bland-Altman plots and the repeatability of each method was assessed.

RESULTS

Diameter estimations obtained using the SLRF method were the least scattered although diameters obtained were approximately 3 pixels greater than those measured manually. The SLRF method was the most repeatable and the Gaussian method less so. The Sobel method was the least consistent owing to frequent misinterpretation of the light reflex as the vessel edge.

CONCLUSION

Of the three automated methods compared, the SLRF method was the most consistent (defined as the method producing diameter estimations with the least scatter) and the most repeatable in measurements of retinal arteriolar diameter. Application of automated methods of retinal vascular analysis may be useful in the assessment of cardiovascular and other diseases.

Quantification and characterisation of arteries in retinal images

A computerised system is presented for the automatic quantification of blood vessel topography in retinal images. This system utilises digital image processing techniques to provide more reliable and comprehensive information for the retinal vascular network. It applies strategies and algorithms for measuring vascular trees and includes methods for locating the centre of a bifurcation, detecting vessel branches, estimating vessel diameter, and calculating angular geometry at a bifurcation. The performance of the system is studied by comparison with manual measurements and by comparing measurements between red-free images and fluorescein images. In general an acceptable degree of accuracy and precision was seen for all measurements, although the system had difficulty dealing with very noisy images and small or especially tortuous blood vessels.

Modelling impact-initiated wave transmission through lung parenchyma in relation to the aetiology of exercise-induced pulmonary haemorrhage

Recently we proposed that exercise-induced pulmonary haemorrhage (EIPH) results from locomotory-impact-induced trauma by impact of the scapula on the chest wall during footfall and the consequent transmission of waves through the lung. A computational model has been developed to demonstrate that wave amplification and focusing occur in the dorsocaudal tip of the lung for waves originating on the anterior subscapular surface. The propagation of an acoustic wave was investigated in a simplified 2-dimensional representation of a vertical anterio-dorsal section of horse lung. It was demonstrated that a complicated pattern of waves is transmitted from the scapula to the dorsal region. Wave motion was characterised using the instantaneous rate of change of pressure with time (dp/dt) which is associated with lung injury. Due to wave reflection and focusing, dp/dt is transiently very high on the spinal and diaphragmatic lung walls, particularly in the vicinity of the dorsal tip. The model therefore predicts that lung injury may occur in the region in which EIPH is reported to originate.

Advanced spectral estimators for detailed blood flow studies

Recent publications have emphasized the relationship between the spectrum of the backscattered acoustic signal, beam geometry, and flow patterns in the measurement of blood flow by Doppler ultrasound. On this basis, we believe that in the future more importance will be placed on analyzing various characteristics of the spectral shape rather than absolute parameters of measurement, such as the mean frequency. The potential of this approach for extracting more information from the raw Doppler signal is introduced by considering the Spectral Broadening Index (SBI). We explain the use of the SBI parameter for measuring flow angle under restricted flow conditions. This is done by using an analytic/computational model for prediction of the spectral broadening effect. By simulation study, the performance of various spectral estimators for determining the SBI from finite Doppler signal segments is evaluated.