Discussion

The purpose of this study was to propose a novel method for semi-automatic segmentation of the boundary of thyroid nodules on non-preprocessed ultrasound images.

Therefore, the correlation and accuracy between AB and MB was an important issue.

The results in Fig. 5-3 and Fig. 5-4 showed that the application of our proposed method on clinical thyroid nodules perfectly conformed to the original design concept, and the boundary drawn by our proposed method agreed well with gold-standard, as shown in Fig. 5-5. Furthermore, our proposed method achieved high effectiveness and successfully prevented the influence from the variation of the nodule’s inner tissues without any preprocessing of ultrasound images. It could handle complicated cases with blurred boundaries, inhomogeneity, or cysts in nodule, as shown in Fig. 5-6.

The overall results of a large amount of cases showed that our proposed method achieved a high degree of correlation between AB and MB with low error rates (summarized in Table 5-2), as demonstrated in Fig. 5-5, Fig. 5-6, and the second and third row of Fig. 5-7. Besides, the results showed that the difference of all the metrics between the training and testing dataset was around 0.35%, and this difference was statistically insignificant proved by Mann-Whitney U test. These results clearly indicated the reliability and consistency of our proposed method.

The results of benign and malignant nodules showed that our proposed method performed ideally both on benign or malignant cases (Table 5-3). The performance metrics of benign cases were slightly better than that of malignant cases, but the difference in-between was around only 0.46% and statistically insignificant, expressing the

reliability and high degree of accuracy on both benign and malignant cases.

Three standardized methods used for comparison were chosen in order to avoid additional adjustments, thereby reinforcing the impartiality and credibility of the comparison. The manipulations to be avoided included additional filtering methods, extra parameters, human intervention, or subjective errors. Consequently, two of the standardized methods, ACM and WM, established with OpenCV library were chosen for comparison as in work by Chang et al.[44]. The third standardized method, DRLSE [45], was chosen in comparison and realized with the source code provided by author(s) [46].

The comparison results with the same dataset were shown in Table 5-4, Table 5-5, and Fig. 5-7. Our proposed method significantly outperformed ACM in all performance metrics and also significantly outperformed DRLSE in most metrics except FPR. As for comparison with WM, the TPR in WM was 1% slightly higher than our proposed method, but was accompanied by the largest FPR (17.50%) due to the boundaries outreaching unnecessary areas, as shown in the fourth row of Fig. 5-7. The FPR in WM was 9.82%, 9.19%, and 9.75% higher than our proposed method, ACM, and DRLSE, respectively.

The NHD in WM was much larger than other methods comparative to NMD, and this implied that there were extreme outliers in WM. Besides, the JSI and PPV of our proposed method were 6.06% and 8.00% significantly higher than that of WM, respectively. WM referred to more information from human intervention, but detected boundary with excessive sensitivity leading to a large error rate, therefore reduced the effectiveness of the boundary segmentation.

Sensitivity analysis showed that the performance varies with different values of w and θThreshold, but not with a and b. It was observed that larger w resulted in worse performance.

This might be due to that a larger value of w might cover too many unnecessary boundary candidate points. By the way, smaller θThreshold was accompanied with worse performance.

Too small value for θThreshold limited the path searching and might cause the path to deviate from the real boundary of nodule, therefore resulting in low TPR. On the other hand, the choice of value for a and b had very limited influence on overall performance. It implied that direction searching method was the principal filtering step of our proposed method, so the two following steps only played an auxiliary role.

Since OpenCV is an optimized cross-platform library [47], the two methods coded with OpenCV, ACM and WM, showed good time performance. However, WM was slightly slower than our proposed method due to the initial requirement of distinguishing the foreground from the background. DRLSE had a poor time performance of 16.01 seconds per nodule. It might be attributed to two reasons: the first was that DRLSE was realized with the version of full domain implementation rather than the version of narrowband implementation, and had higher computational cost [45]. The second reason was that the author(s) provided DRLSE source code in MATLAB, different from the Microsoft Visual Studio C# we used to code our algorithm and perform the comparison study. The transition efficiency between MATLAB and the Microsoft Visual Studio C#

had been reported as a problem [48], [49].

Speckle poses a challenge to computer-aided systems for ultrasound images but also provides useful information about tissue characteristics [9]-[14], [15]-[16]. Koundal et al.

applied Gaussian filter in clustering and contour evolution to reduce the impact of speckle [14]. Similarly, Savelonas et al. utilized local binary pattern operator which represented regional textural properties instead of single pixel [12]. These approaches averaged out speckle but meanwhile erased the information carried by the speckles. On the contrary, our method did not smooth out the speckles in every step of the algorithm, and, therefore, the speckle information was kept and was also an integral part of our proposed segmentation algorithm.

The gold-standard for performance validation was defined by a single experienced radiologist. The reason for not using multiple radiologists was that unlike other medical decisions it is not straightforward to pool together different boundaries delineated by different radiologists. Averaging different manually delineated boundaries may not be appropriate because the discrepancies are usually due to different viewpoints about the same nodule image instead of the observation errors. For example, for lobulated nodules radiologists usually have different viewpoints on whether a certain adjacent mass should be included as part of the nodule. To avoid such issue, we used the boundary manually delineated by a single experienced radiologist as the gold-standard in this study.

Nevertheless, the gold-standard boundary established by only one experienced radiologist was still a limitation of this study.