Introduction
Material and methods
Multi-layer segmentation
Score | Relevant | Feature |
---|---|---|
51/51 | Yes | Contrast of the region R [9] |
51/51 | Yes | Mean square error between the intensity profile I of the region R and the linear interpolation of R as a polynomial of degree 1, calculated as \(\text {Average}\big (|I(i)-P(i)|^2\big )\), where P is the 1-degree interpolation of I, and i is the index of the \(i\mathrm{th}\) pixel along the A-line |
50/51 | Yes | Homogeneity of the region R [9] |
50/51 | Yes | Gradient \(\textsc {g}_3\) at the MA interface |
49/51 | Yes | Correlation of the region R [9] |
48/51 | Yes | Gradient \(\textsc {g}_2\) at the IM interface |
34/51 | Yes | Gradient \(\textsc {g}_1\) at the LI interface |
34/51 | Yes | Monotony index, defined as the median absolute distance of all piece-wise permutations required to sort the intensity values of the region R in a monotonically decreasing fashion, calculated as \(\text {Median}(|i-i'|)\), where \(i'\) is the new index of i after the A-line has been sorted in a descending intensity order |
22/51 | No | Sum of all negative gradient values \(I_G\) in the region R, calculated as \(\sum _i\big (min(I_G(i), 0\big )\)
|
21/51 | No | Entropy of the region R [9] |
13/51 | No | Median image intensity in the media layer, between the IM and MA interfaces |
12/51 | No | Energy of the region R [9] |
5/51 | No | Median image intensity in the adventitia layer, between the MA and AP interfaces |
4/51 | No | Median image intensity in the intima layer, between the LI and IM interfaces |
4/51 | No | Distance between the LI and IM interfaces |
3/51 | No | Distance between the LI and AP interfaces |
1/51 | No | Distance between the LI and MA interfaces |
Healthy region classification
Rationale
Feature selection
healthy
or diseased
label.1 Boruta was applied to identify the important features. The maximum number of iterations was 1000. The Bonferroni correction was applied. The value \(p<0.01\) was considered to indicate a statistically significant difference. This experiment was conducted 51 times (odd number), each time with a different subset of randomly gathered A-lines from the training set. Finally, relevant features were defined as those that were labeled as “important” by Boruta at least 26 times (i.e., for more than half of the 51 total number of experiments).Classification
healthy
or diseased
label is determined by majority voting between the corresponding column in the three frames. The aim of this operation is to exploit the consistency of the data along the axis of the pullback to cope with frames that may be hindered by image noise.St Jude Medical Erasmus MC Rotterdam, The Netherlands | Terumo University Hospital of Auvergne Clermont-Ferrand, France | |
---|---|---|
Pullback speed (mm/s) | 20 | 20 |
Acquisition rate (fps) | 100 | 158 |
Acquisition length (mm) | 54 | 80 |
Number of frames | 271 | 632 |
Resolution (axial/lateral) (\(\upmu \text {m}\)) | 20/30 | 20/15 |
Depth of the scan range (mm) | 4.3 | 4.5 |
Dimensions of the polar image (pixels) | 968 \(\times \) 504 | 512 \(\times \) 512 |
Pixel size (\(\upmu \text {m}\)) | 4.5 | 8.8 |
Data collection
Image analysis procedure
Image selection
Manual reference annotations
Method evaluation
Results
Parameter settings
Contour segmentation
Training set | Testing set | |||||
---|---|---|---|---|---|---|
All | SJM-OCT | Terumo-OCT | All | SJM-OCT | Terumo-OCT | |
Intima–Media | ||||||
Method vs
\(\mathcal {A}_1\)
| 25 ± 37 | 28 ± 44 | 23 ± 32 |
\(\varvec{29 \pm 46}\)
| 33 ± 48 | 27 ± 44 |
Inter-analysts | 21 ± 25 | 21 ± 28 | 21 ± 22 | 20 ± 34 | 24 ± 45 | 18 ± 24 |
Intra-analyst | 23 ± 40 | 31 ± 54 | 16 ± 20 | 15 ± 21 | 15 ± 18 | 15 ± 23 |
Media–adventitia | ||||||
Method vs
\(\mathcal {A}_1\)
| 27 ± 42 | 29 ± 45 | 25 ± 39 |
\(\varvec{30 \pm 50}\)
| 31 ± 49 | 29 ± 50 |
Inter-analysts | 20 ± 23 | 22 ± 28 | 19 ± 18 | 23 ± 48 | 27 ± 69 | 20 ± 25 |
Intra-analyst | 20 ± 37 | 28 ± 52 | 14 ± 13 | 17 ± 28 | 20 ± 38 | 15 ± 20 |
Adventitia–tissues | ||||||
Method vs
\(\mathcal {A}_1\)
| 37 ± 48 | 38 ± 49 | 37 ± 47 |
\(\varvec{50 \pm 64}\)
| 50 ± 66 | 49 ± 62 |
Inter-analysts | 25 ± 28 | 27 ± 32 | 24 ± 25 | 32 ± 53 | 37 ± 72 | 29 ± 34 |
Intra-analyst | 24 ± 40 | 34 ± 55 | 16 ± 20 | 24 ± 37 | 27 ± 47 | 22 ± 28 |
Healthy region classification
healthy
and diseased
labels resulting from the automatic classification method were in good accordance with the manual annotations, as displayed in Fig. 7. For all the 260 analyzed images, the median value (and interquartile range) of the accuracy, sensitivity, and specificity was 0.91 (\([0.75 - 0.98]\)), 0.92 (\([0.71 - 1.00]\)), and 1.00 (\([0.92 - 1.00]\)), respectively. The median Dice coefficient was 0.93, with an interquartile range of \([0.78 - 0.98]\) and an average (±SD) value of \(0.83 \pm 0.25\). More specifically, the median, interquartile range, and average was 0.92, \([0.74 - 0.98]\), and \(0.79 \pm 0.30\) for the SJM-OCT images, and 0.92, \([0.81 - 0.98]\), and \(0.86 \pm 0.19\) for the Terumo-OCT images, respectively. For each of the 260 images of the testing set, the average percentage of healthy regions compared to the entire circumference of the wall, as annotated by the analyst \(\mathcal {A}_1\), was \(69 \pm 31\%\). This is to be compared with the corresponding ratio derived from the automatic method, which was \(61 \pm 29\%\). To assess the usefulness of feature selection, these results were confronted to those obtained when using the full set of 17 features. Similar results were systematically found: the median value (and interquartile range) of the Dice coefficient, accuracy, sensitivity, and specificity were 0.92 (\([0.79 - 0.99]\)), 0.89 (\([0.76 - 0.98]\)), 0.93 (\([0.69 - 1.00]\)), 1.00 (\([0.93 - 1.00]\)), respectively. The healthy region classification method was finally applied to the first 200 frames of a single pullback as a proof of concept to automatically highlight healthy regions, as shown in Fig. 8.
Computational speed
Discussion and conclusions
On the contour segmentation method
On the healthy region classification method
healthy
are trustworthy, and only regions labeled as diseased
—that must be analyzed anyway—should be visually inspected to confirm the classification result. Let us also note that the accuracy (Dice coefficient, accuracy, sensitivity and specificity) is very close to the inter-observer variability (Fig. 7), which contributes to validate the performances of the method.healthy
or diseased
label is applied to different regions of the image. In contrast to several previous studies [1, 20, 21, 24, 26], the present method is not capable of characterizing different tissue types, such as lipid, fibrous, or calcium. To address this limitation, future work will focus on a cascade approach, where the artery is first roughly partitioned in healthy and diseased regions using the present framework, before a second classifier based on different image features is applied only to diseased regions for finer tissue and plaque type characterization. However, the present method can potentially be used as a preprocessing step to guide existing classification approaches in diseased regions.