Non-invasive diagnosis of deep vein thrombosis from ultrasound imaging with machine learning

Abstract

Deep vein thrombosis (DVT) is a blood clot most commonly found in the leg, which can lead to fatal pulmonary embolism (PE). Compression ultrasound of the legs is the diagnostic gold standard, leading to a definitive diagnosis. However, many patients with possible symptoms are not found to have a DVT, resulting in long referral waiting times for patients and a large clinical burden for specialists. Thus, diagnosis at the point of care by non-specialists is desired. We collect images in a pre-clinical study and investigate a deep learning approach for the automatic interpretation of compression ultrasound images. Our method provides guidance for free-hand ultrasound and aids non-specialists in detecting DVT. We train a deep learning algorithm on ultrasound videos from 255 volunteers and evaluate on a sample size of 53 prospectively enrolled patients from an NHS DVT diagnostic clinic and 30 prospectively enrolled patients from a German DVT clinic. Algorithmic DVT diagnosis performance results in a sensitivity within a 95% CI range of (0.82, 0.94), specificity of (0.70, 0.82), a positive predictive value of (0.65, 0.89), and a negative predictive value of (0.99, 1.00) when compared to the clinical gold standard. To assess the potential benefits of this technology in healthcare we evaluate the entire clinical DVT decision algorithm and provide cost analysis when integrating our approach into diagnostic pathways for DVT. Our approach is estimated to generate a positive net monetary benefit at costs up to £72 to £175 per software-supported examination, assuming a willingness to pay of £20,000/QALY.

Conflict of interest statement

B.K., M.H., R.M., and N.C. are scientific advisors for ThinkSono Ltd. B.K. is also advisor for Ultromics Ltd and Cydar medical Ltd. A.M. was an employee of ThinkSono Ltd until September 2020. J.O., F.A-.N., and S.M. are employees of ThinkSono Ltd, M.D.S. and A.C.R. acted as contractor for ThinkSono Ltd. All authors had full access to all data during this study and accept responsibility to submit for publication. B.K., A.M., F.A-.N., and S.M. are joint inventors on a patent held by ThinkSono Ltd. The remaining authors declare no competing interests. The views expressed are those of the author(s) and not necessarily those of ThinkSono Ltd, the NHS, the NIHR or the Department of Health. The remaining authors declare no competing interests.

© 2021. The Author(s).

Figures

Fig. 1. Consort diagram for study enrolment.

Allocation, and analysis in External Validation Set 1 (EVS1).

Fig. 2. Qualitative example images for our model’s segmentation performance.

The segmentation is robust throughout compressions. The vein area is evaluated for complete compressibility to exclude DVT. Device: Clarius L7 (2017).

Fig. 3. Evaluation results for ESV1 on patient level.

Receiver operator characteristics on EVS1 resulting from fourfold cross validation (a). Confusion matrices are shown in (b) for the optimal threshold (* in (a)) in each fold. Frame colours in (b) correspond to ROC fold colours in (a)). Vessel status is extracted automatically from 53 patients in EVS1 through the combination of fold-specific groin and knee model pairs.

Fig. 4. Evaluation results for ESV1 on compression sequence level.

Receiver operator characteristics for the correct compression classification per ultrasound sequence/anatomical landmark on EVS1 (a). Confusion matrices (b) at optimal thresholds (* in (a)) per fold. Frame colours correspond to ROC fold colours in (a). Vessel status is extracted automatically through the ML models from the 121 available anatomical landmark sequences in EVS1.

Fig. 5. Confusion matrices for EVS2 for each fold.

Frame colours correspond to cross-validation folds (a–d) from Figs. 3 and 4. Vessel status is extracted automatically from all 30 patients in EVS2. Note that this experiment comprises of four DVT positive and 26 DVT negative patients, thus sensitivity/true positive rate is discretised in a ROC curve with only four steps, which makes a plot less meaningful.

Fig. 6. Costs of the guidance tool vs. net monetary benefit (NMB) per examination when implementing ML-guided DVT diagnostics into clinical diagnostic pathways.

The NMB has been simulated with a deterministic model for each of the diagnostic algorithm variants in Fig. 11 at the mean (solid line) and the 95 CI interval (shaded area) from Table 5 to show possible optimistic and pessimistic scenarios. The red lines on the y-axis mark the maximal attainable NMB range when examination costs are zero.

Fig. 7. Consort diagram for inclusion of volunteer scans into the training set and internal validation set.

Dataset curation for the training and internal validation data. Our approach can be trained from image data that originates predominantly from healthy volunteers.

Fig. 8. Examples of the chosen anatomically salient landmarks and overview over the investigated anatomy.

Images have been acquired by different acquisition devices and from different subjects. This figure illustrates the diversity in our dataset. See the overview above the table and Tables 7 and 8 for a description for the location of these landmarks. These example images have been manually cropped and contrast normalised for better readability.

Fig. 9. Overview over the AutoDVT prototype core algorithm.

a whole overview and b overview over the individual branches. A U-Net serves as a backbone for automatic delineation of vein and arteries (b). The prediction of the anatomical location of the image is based on our previous work. Network branches predict the anatomical location and whether the vessel is open or closed under pressure. Landmark predictions are performed from the learned numeric representation in the bottleneck layer; vessel compression state is predicted from the output segmentation mask. The network components are connected and can be trained through back-propagation in an end-to-end manner. The input is a stack of nine images (individual video frame images resampled to 150 × 150 pixels) from an ultrasound video stream that moves by one in a sliding window fashion. A single segmentation mask is produced for the last-most image within approximately 25 ms. Two separate models with identical architecture are trained, one for the groin area (LM0–LM5) and one for the knee area (LM8–LM10). Each model holds 31,475,527 parameters. (OC = open/close).

Fig. 10. Prototype implementation user interface.

The AutoDVT software instructs users to locate a given landmark, instructs to perform a correct compression and evaluates the result automatically.

Fig. 11. Possible integration strategies for our approach into DVT diagnostics pathways.

a current clinical algorithm to diagnose DVT without software support according to UK NICE guidelines and b–f possible variants to integrate ML software support into the clinical pathway. Algorithms 1–3 shown in (b–d) generate a positive net monetary benefit (cf. Fig. 6). The examined modifications have been suggested by health economics and clinical experts. Note that treatment options may further depend on the age of the clot, which might be manually estimated during confirmatory ultrasound scans.