4763

Automatic detection of corrupted frames in cardiac DTI with machine learning

Francesca Cavallo¹, Pedro Ferreira¹, Zohya Khalique¹, Andrew Scott¹, Sonia Nielles-Vallespin², and David Firmin¹

¹Cardiovascular BRU, Royal Brompton Hospital, London, United Kingdom, ²NHLBI, National Institutes of Health, MD, United States

Synopsis

In vivo cardiac DTI is capable of probing the microstructure of the myocardium and its dynamics throughout the cardiac cycle. The typical cardiac DTI scan data will contain corrupted frames due to cardiac and respiratory motion. Currently an experienced observer identifies corrupted frames by means of a visual assessment and manually removes them. In this work we show that machine learning can be used to accurately assess DTI corrupted frames, reducing the user input, accelerating analysis and removing human subjectivity.

Purpose

The objective of this work is to accelerate the post-processing of cardiac diffusion tensor imaging (DTI) by automating the removal of corrupted frames with machine learning.

Background

In vivo cardiac DTI is capable of probing the microstructure of the myocardium and its dynamics throughout the cardiac cycle. A typical DTI study of one mid-ventricular slice scanned at two stages of the cardiac cycle will typically contain more than 180 frames, with approximately 14% of them corrupted by signal loss due to cardiac and respiratory motion. Currently an experienced observer identifies corrupted frames by means of a visual assessment and manually removes them, before the diffusion tensor is calculated. This can take up to 5 minutes per subject.

Methods

Data from a total of 51 subjects (19 healthy, 32 patients) scanned at systole and diastole were used to assess the accuracy of machine learning (ML). A visual assessment by two experienced users identified 1407 corrupted frames (figure 1) out of a total of 9651 frames (8% in healthy, 18% in disease). Half of the subjects in both cohorts were randomly selected for machine training. The remaining data were used to determine the accuracy of the algorithm.

A ML algorithm was used, where images are converted to a vector quantization of affine invariant descriptors¹ before using a quadratic support vector machine classifier (MATLAB, Mathworks, Massachusetts, USA). The accuracy of corrupt frame identification by machine learning was compared to the best intrasubject image similarity algorithm we tested, a Pearson correlation coefficient (PCC) test (with the optimum threshold for this data).

Three common cardiac DTI parameters: Fractional Anisotropy (FA), Mean Diffusivity (MD), and Helix-angle range (HAr); were compared for the same test cohorts when processing the data with the corrupted frames removed by the initial visual assessment, and by the two automated methods: ML and PCC.

Results

The accuracy of corrupt frame identification with ML was statistically significantly higher than that of the PCC tests (p = 0.002, Wilcoxon signed rank). The overall confusion matrices are shown in figure 2. An intersubject mean+/-SD accuracy of 95+/-5% was achieved with ML, above the mean 92+/-11% accuracy of PCC. Machine learning yielded significantly lower false negatives (figure 3) and a lower inter-subject accuracy variance.

An overall average mean difference to the visual assessment of 6% and 8% are observed for ML and PCC respectively when combining all DTI parameters (figure 4). The ML algorithm takes approximately 25 seconds to assess the entire database of 51 subjects.

Discussion & Conclusions

ML can be used to accurately assess DTI corrupted frames with better efficiency than image similarity measures. Therefore, ML is capable of reducing the user input, accelerating analysis and removing human subjectivity. As the DTI patient database grows, more data will be available for training, potentially improving accuracy further.

Acknowledgements

This project was supported by the NIHR cardiovascular Biomedical Research Unit at the Royal Brompton and Harefield NHS Foundation Trust and by the Wellcome Trust [WHCC NCP142].

References

1 – Csurka, G., Dance, C., Fan, L., et al. Visual categorization with bags of keypoints. (2004): 1--2.

Figures

Figure 1: Example of a good and a corrupted DTI frame

Figure 2: Confusion matrices for machine learning and Pearson correlation coefficient test.

Figure 3: Example of DTI frames for a healthy volunteer in diastole. The red frames are the ones identified as bad by two observers. ML – frames identified as bad with machine learning. PC – frames identified as bad with the PCC test (reference frame marked as REF).

Figure 4: Absolute mean difference comparison of DTI parameters (FA – fractional anisotropy, MD – mean diffusivity, HAr – Helix-angle range) between the original visual assessment and the two automated methods of removing corrupted frames (ML and PCC).

Proc. Intl. Soc. Mag. Reson. Med. 25 (2017)

4763