5 курс / Онкология / Возможности_систем_автоматического_анализа_цифровых_рентгенологических

69

Based on the results of the interpretation analysis with cases pulmonary nodules and masses from radiologists, it was revealed that the major difficulties were caused by Case N4 (Figure 26) with the localization of a solid nodule in the apical section of the upper left pulmonary lobe according to CT and visualized behind the shadow of the collarbone on a PA chest X-ray. The percentage of erroneous interpretations of the X- ray images in this case was 48.4%, while 56.2% of radiologists with experience in thoracic radiology made a mistake in interpreting this case, while among the doctors without this experience, 45.9% of specialists made a mistake (Table 12). A greater number of radiologists who misinterpreted the digital X-ray image with case N4 were among specialists with more than 10 years of experience and amounted to 54.9% versus 46.8% of radiologists with less than 10 years of experience.

Figure 26: X-ray examination data, Case N4, patient with adenocarcinoma in C1+2 of the left lung

Along with this, X-rays with pathology localization behind the shadow of the 1st rib and partially behind the mediastinal shadow, as in Case N1 and N2, caused difficulties; the percentage of wrong answers among radiologists was 43.8% and 46.5%, respectively. At the same time, the largest number of missing pathologies on digital X- rays with these two cases were among the doctors with more than ten years of experience: 56.9% of doctors made a mistake when analyzing Case N1 and 58.8% of doctors – when analyzing Case N2.

The highest percentage of correct answers was obtained in the analysis of digital X-ray images with solid structure localized in the intercostal space, i.e., Cases N5 and

71

CHAPTER 4 COMPARATIVE ANALYSIS OF POSSIBILITIES OF VARIOUS AUTOMATED ANALYSIS SYSTEMS OF X-RAYS IN PATIENTS WITH SUSPECTED NODULES AND MASSES

An analytical validation of four programs (A, B, C, D) has been completed to research the possible use of automated analysis systems of PA chest X-rays to detect pulmonary nodules and masses. for the analysis of chest X-rays, which bill themselves as systems capable of automatically detecting and marking pathology on PA chest X- rays.

Analytical validation was done with three sampling packages containing various pathology occurrence:

Sampling Package 1 (5,150 X-rays: 150 (3%) X-rays with various pathologies manifested as nodules and masses)

Sampling Package 2 (100 X-rays: 6 (6%) X-rays with various pathologies manifested as nodules and masses)

Sampling Package 3 (300 X-rays: 150 (50%) X-rays with various pathologies manifested as nodules and masses)

The diagnostics efficiency values of Program A are given in Table 13, Figure 27.

Table 13: Diagnostics efficiency values of Program A when detecting pulmonary

nodules and masses on survey PA chest X-rays

73

(0.989). The sensitivity value was higher (83.3%); whereat, the positive predictive value amounted to 0.833. As with the analysis of Sampling Package 1, we obtained high likelihood ratio of a positive test (78.3).

The next stage of testing Program A was to test its diagnostics efficiency with Sampling Package 3; pathology distribution of this package has made 50%. Based on the research results for this stage, the program did not pass the threshold for admission to clinical validation, having obtained AUC of 0.770 [44]. However, the specificity value remains high (99%). The sensitivity value turned out to be low (55%).

This may serve as a reason to conclude that using this program at a diagnostic pulmonary center where patients with pulmonary pathology dominate, will not be effective. At the same time, the possibility to pass clinical validation aimed to use the program for chest screening examinations is worth noting [44].

Following the more detailed analysis of the results obtained during the testing of Program A, 40% of all TB cases and 43% of all lung cancer cases were missed in Sampling Package 3.

Program B was also tested with all three packages (Table 14, Figure 28).

Table 14: Diagnostics efficiency values of Program B when detecting pulmonary nodules and masses on survey PA chest X-rays

75

Under the next analyzed Sampling Package 3 with the distribution value of 50%, Program B, as at the first two stages, was close to the threshold value, but did not manage to overcome it (AUC of 0.770) [44]. Retention of high specificity (100%) and the increase of sensitivity up to 54% have been noted.

Speaking of Program B, it should be noted that, despite the high specificity values (91-100%) in all three packages, not enough AUC value (0.723-0.770) necessary to pass the threshold for further clinical validation is a significant reason for the manufacturer to continue improving the program; admittance of Program B to further testing (clinical validation) is currently not feasible.

Program B missed 44% of all TB cases and 42% of lung cancer cases [44].

Due to the access limitations, Programs C and D were tested only with Sampling Packages 2 and 3.

The diagnostics efficiency values of the program are given in Table 15, Figure

29.

Table 15: Diagnostics efficiency values of Program C when detecting pulmonary nodules and masses on survey PA chest X-rays

77

Figure 30: ROC-curves based on the analysis of Sampling Package 2 (a) and Sampling Package 3 (b) by Program D

The AUC value from the analysis of Sampling Package 2 with Program D made up to 0.787; this can serve as a reason to recommend the program for the further work on the system to overcome the threshold and gain access to clinical validation. The specificity value of Program D reached 90%; in general, this can be compared with the values of the previous programs reviewed, as well as its negative predictive value of 0.977. The sensitivity value obtained during the testing of Program D with Sampling Package 2 has made 66.7%; this correlates with the positive predictive value of 0.308.

The results of testing Program D with Sampling Package 3 with an indicator of pathology distribution of 50% appeared to be better. The sensitivity value reached 87%, becoming the highest value for all programs at this stage of the research, as well as the