Introduction
Material and methods
Data acquisition
Image preprocessing and indicators extraction
Image preprocessing and vessel segmentation
Image indicator extraction
Results
Dataset generation
Type of cancer | Histopathology | Patients | Images | Total |
---|---|---|---|---|
Benign | Cyst | 3 | 150 | 20 patients 890 images |
Polyp | 4 | 130 | ||
Reinke’s edema | 5 | 250 | ||
Papilloma | 5 | 230 | ||
Dysplasia mild | 3 | 130 | ||
Malignant | Dysplasia severe | 4 | 130 | 11 patients 465 images |
Carcinoma in situ | 4 | 155 | ||
Carcinoma | 3 | 180 | ||
Total | 31 | 1355 | – |
Dataset based on the degree of disorder of vascular patterns
Dataset based on the histopathologies of the larynx
-
Dataset II CE + NBI images of the benign histopathologies. 20 patients with 890 images labeled into four groups: cyst, polyp & reinke’s edema, papilloma, and dysplasia mild.
-
Dataset III 465 CE + NBI images belonging to 11 patients diagnosed with malignant histopathologies labeled into three groups: dysplasia severe, carcinoma in situ and carcinoma.
-
Dataset IV CE + NBI images belonging to 31 patients with benign and malignant histopathologies that included a total of 1355 images labeled into two groups: benign and malignant.
Qualitative analysis
-
The HGD indicator shows changes in the energy concentration with respect to the angle of the gradient vectors. Parallel vessel patterns show energy concentration of the gradient vector in two angles, while more chaotic vessel structures show a leakage in the energy distribution and even an equal distribution of energies (flat indicator) in the presence of spiral vessel patterns. It is possible to assume that the energy and energy-related characteristics of the HGD indicator can differentiate between different vascular patterns.
-
The matrix of row averages for each rotation angle of the RIA indicator displays highly concentrated energies in two rotational angles when vessel patterns are parallel. This produces a final RIA containing a few number of main peaks of high amplitude. The more the vessel patterns become chaotic, the quantity of peaks and the energy leakage increase in the RIA. Energy-related features can therefore be used for characterizing vessel patterns.
-
The displayed signal for both ANG and DIS indicators (Fig. 6) are a concatenated version for several vessel segments. This is why some signal discontinuities can be observed in the indicators. Disrespecting these discontinuities, we can observe that a vessel with significant curve patterns produces ANG and DIS indicators involving an increased number of changes per distance unit. This means that the quantity of changes of sign in their derivatives and the polynomial fitting errors will be higher for disorder patterns than for ordered ones, making it more suitable for distinguishing between patterns.
-
CUR indicator variance increases when the vessel patterns become disorder. This is mainly because disorder patterns involve a higher number of loops and therefore more significant curvature’s values. For this indicator the features are also based on energies and peaks in the signal and also statistical values as variance.
Features classification performances
Database | Accuracy | Sensitivity | Specificity | AUC |
---|---|---|---|---|
Dataset I | 0.973 | 0.980 | 0.983 | 0.977 |
Dataset II | 0.846 | 0.819 | 0.942 | 0.917 |
Dataset III | 0.864 | 0.856 | 0.931 | 0.917 |
Dataset IV | 0.847 | 0.806 | 0.868 | 0.837 |
Database | Accuracy | Sensitivity | Specificity | AUC |
---|---|---|---|---|
Dataset I | 0.968 | 0.976 | 0.978 | 0.973 |
Dataset II | 0.816 | 0.757 | 0.926 | 0.901 |
Dataset III | 0.873 | 0.864 | 0.931 | 0.921 |
Dataset IV | 0.837 | 0.834 | 0.839 | 0.837 |
Database | Accuracy | Sensitivity | Specificity | AUC |
---|---|---|---|---|
Dataset I | 0.965 | 0.974 | 0.978 | 0.989 |
Dataset II | 0.892 | 0.879 | 0.958 | 0.969 |
Dataset III | 0.877 | 0.873 | 0.939 | 0.956 |
Dataset IV | 0.912 | 0.871 | 0.933 | 0.953 |
Database | Accuracy | Sensitivity | Specificity | AUC |
---|---|---|---|---|
Dataset I | 0.966 | 0.975 | 0.979 | 0.996 |
Dataset II | 0.906 | 0.900 | 0.965 | 0.981 |
Dataset III | 0.884 | 0.879 | 0.943 | 0.973 |
Dataset IV | 0.911 | 0.939 | 0.858 | 0.979 |