SCI Rep: Uses computed tomography and deep neural networks to automatically detect enlarged extraocular muscles in Graves' eye disease

Background: Graves' ophthalmopathy (GO) is a chronic autoimmune disease that affects the posterior bulbar tissue and extraocular muscles and has a strong etiological link
to autoimmune thyroid disease.
In actual clinical practice, approximately 40%–60% of GO patients have been reported to have extraocular muscle dysfunction and have a significant negative impact
on quality of life.
Early detection of extraocular muscle abnormalities and orbital imaging may be necessary for successful management of thyroid myopathy
.
In practice, orbital imaging
is unlikely unless the patient complains of diplopia.
In addition, radiologists may not always be able to interpret these findings, especially in areas where there is a shortage of
doctors.
In some parts of the developing world, there may be fewer
facilities for adequate imaging than radiologists.
Supervised machine learning systems known as neural networks have been applied to medical research
.

Many studies have been conducted
on the diagnostic and classification performance of deep learning (DL) systems using CT images.
However, to our knowledge, there have been no reports of DL systems using CT images to classify magnified extraocular (EEM) images of GO patients and normal extraocular muscle (NEM) images in normal subjects
.

Objective: The objective of this study is to develop a diagnostic software system in which a DL system can evaluate EEG and orbital CT images
in patients.

Methods: Deep neural network was used to perform orbital coronary CT examination of the extraocular muscles of 371 patients with Graves' ophthalmopathy (GO) (199 GO patients and 172 normal extraocular muscle patients), and a diagnostic system
of extraocular muscle deep neural network was established.
Patients are classified as EEM with
GO when at least one rectus muscle (right or upper left, lower, medial, or lateral) is greater than or equal to 4.
0 mm.
We used 222 patient data as training data, 74 images as validation test data, and 75 images as test data to train the deep neural network to determine the thickness
of the extraocular muscles on computed tomography.
We then verified the performance of
the network.

Results: We used EEM images from 199 GO patients (56 men and 143 women) (mean age 55.
9±13.
7 years) and from 172 controls (40 men and 132 women; The mean age was 52.
6 ±18.
4 years).

We found no significant differences in age (p = 0.
21) or sex (p = 0.
85) between the two groups (Table 1).

Table 2 shows the right and left upper, lower, medial, and lateral rectus muscles
in both groups.
All right or left rectus muscle thickness differed significantly between groups (p < 0.
001).

In the test data, the area under the curve (AUC) for neural network diagnostics was 0.
946 (95% confidence interval [CI] 0.
894–0.
998), and the subject operating characteristics (ROC) analysis showed a sensitivity of 92.
5% (95% CI 0.
796–0.
984) and a specificity of 88.
6% (95% CI 0.
733–0.
968) (Figure 1).

For the test data, it took 276.
2 seconds to analyze the CT scans (3.
6 seconds/patient)
of 75 patients.

Table 1 Difference
between the maximum diameter of extraocular hypertrophy (EEM) and normal extraocular muscle (NEM).
Unless otherwise noted, EEM and NEM data are expressed as mean ± standard deviation
.
EEM enlarges the extraocular muscle, IRM inferior rectus, LRM external rectus muscle, MRM internal rectus muscle, NEM normal extraocular muscle, SRM superior rectus muscle

Figure 1 (a) Receiver operating characteristics (ROC) curve
for validation data.
The area under the curve (AUC) for neural network diagnostics was 0.
953, and ROC analysis showed 89.
7% sensitivity and 94.
3% specificity
.
(b) ROC curve
of the test data.
The AUC for neural network diagnostics was 0.
946, and ROC analysis showed 92.
5% sensitivity and 88.
6% specificity
.

Figure 2 Computed tomography (CT) slice images (a) and heat maps (b)
of healthy participants.
CT slide images (c) and heat maps (d)
of patients with Graves' eye disease.
Blue indicates the intensity
of attention of deep neural networks.
On orbital coronal CT images, the color intensity is higher
in the rectus area.
The deep neural network classified the extraocular muscles of Graves' ophthalmopathy patients as enlarged, while the control group classified the extraocular muscles as normal, with an emphasis on the rectus muscle
.

Figure 3 Coronal scan in the nearaxial plane at 90° to the orbital axis is reconstructed by axial scanning (a).

In coronal scans, six consecutive slices (2 mm thick) from the posterior edge of the eye to the orbital tip are used (b).

Figure 4 Coronal sections (a) and results (b) for segmentation of the
eyeball.
Coronal sections (c) and results (d) are used for orbital segmentation
.
Coronal slices (e) and region of interest (area within the blue square) (f) used when the Residual Network-50 identifies the post-bulbous region from (b)
and (d).

Conclusion: Deep learning system using deep neural network can detect EEM
in GO patients.

Original source:

Hanai K, Tabuchi H, Nagasato D, et al.
Automated detection of enlarged extraocular muscle in Graves' ophthalmopathy with computed tomography and deep neural network.
Sci Rep 2022 Sep 26; 12(1)