PROJECT modified final of deserdation.pptx

Project

1. Atlas-based Auto-Segmentation in Head and Neck Cancer Radiotherapy: Clinical Benefits and Challenges Presented by Dayawoti pegu resident MEDICAL PHYSICIST, HBCH&RC VIZAG Guided BY M anil kumar medical physicist, HBCH&RC VIZAG

2. OUTLINES Introduction Atlas based auto segmentation Importance of auto-segmentation Aim of the project Material and Method Results and discussion conclusion

3. Introduction Radiotherapy treatment planning is a time consuming process. The target volumes and organs-at-risks(OAR) are manually delineated for treatment plan generation. In head and neck cancers, accurate contour delineation is essential for a good treatment outcome. To improve contouring efficiency and reduce potential inter-observer variation, the atlas-based auto-segmentation function was introduced.

4. ATLAS BASED AUTO SEGMENTATION An atlas is a predefined volume containing a set of predefined structures. An atlas set contains several different atlases that all have a particular area in common. Atlas-based auto segmentation is used to automatically contour target tumors and normal tissues on the CT images of a new patient using predefined atlases and a non rigid registration technique. Atlas based auto contouring is two types 1. Single atlas set 2. Multi atlas The single atlas based procedure has the disadvantage of selecting only one atlas that is similar to the patient. It is not practically possible to register and choose an atlas that roughly matches the patient with respect to all regions. Because of which the auto-contouring accuracy becomes low. In multi- atlas based procedure, the auto-contouring accuracy will be more, because the most similar structure can be selected from multiple structures. Velocity - single atlas based auto-segmentation SmartAdapt - multi atlas based auto-segmentation.

5. Importance of auto segmentation Interobserver variations. Substantial time. OAR’S may not be routinely contoured. The development of computational tools to automatically generate OAR contours can reduce the time and effort required for HNC contouring.

6. Literature review on inter-observer variations OAR’S DSC MDA(mm) HD(mm) Brainstem 0.88(0.61-0.92) 1.5(1.1-4) 4(2.3-15) Spinal cord 0.78(0.56-0.90) 2.2(0.8-10.4) 12.1(1.7-72.1) Mandible 0.90(0.79-0.94) 1.1(0.8-8.3) 3.4(1.5-38.7) Oral cavity 0.77(0.45-0.91) 4.6(1.8-11.6) 14.5(4.3-30.1) Parotid R .83(0.51-0.90) 2(1.4-4.9) 5.1(3.2-19.2) Parotid L 0.82(0.62-0.88) 1.9(1.2-4.2) 4.9(3.1-16.5) Optic nerve R 0.59(0.55-0.64) 1.1(0.7-1.3) 6.5(4.1-7.7) Optic nerve L 0.59(0.52-0.63) 1.0(0.9-1.4) 6.9(5.1-8.2) Ebbe Laugaard Lorenzen et all. https://doi.org/10.1080/0284186X.2021.1975813 https://doi.org/10.1186/s13014-020-01677-2

7. Aim The aim of the project is to compare the accuracy of Atlas-based auto-segmentation of two applications, i.e., SmartAdapt and Velocity, at the head and neck sites. To analyze dosimetric impact of auto-segmented contours against manually segmented clinical contours.

8. methodology 85 hypopharynx Patients (till Nov 2023). 50 patients selected for creating atlas and 12 patients kept for validation. Patient Selection criteria for creating an atlas 1. Age 30 to 70 years 2. Distance between chin to sternum notch is 7 to 10 cm 5. BMI from 15 to 25 In smartAdapt , selected top five patients with highest similarity for segmentation.

9. In velocity provides tools to improve the atlas based contouring 1. Model based segmentation Brainstem Refined Spinal cord Refined 2. Local deformable registration Brainstem Shaped Spinal cord Shaped On comparison, Procedure 2 was found to give better segmentation accuracy and hence was used for comparison with smartAdapt ® The auto-segmented contours were taken for dosimetric evaluation. Time taken for Manual and the two auto-contouring systems were compared.

10. Local deformable registration in velocity

11. OAR’S for Auto segmentation hypopharynx 1. Brainstem 2. Spinal cord 3. Oral cavity 4. Eye R 5. Eye L 6. Parotid R 7. Parotid L 8. Mandible 9. Optic nerve R 10. Optic nerve L 11.Esophagus

12. PARAMETERS FOR EVALUATION Dice similarity coefficient( DSC) Hausdorff distance (HD) Relative volume difference ( RVD) Mean distance to agreement( MDA) Acceptance criteria according to ‘AAPM TG-132 report’ Geometric parameters Tolerance DSC ~0.80–0.90 HD ~2–3 mm MDA ~2–3 mm RVD ~0

13. The DSC method, calculates the overlapping results of two different volumes according to the equation dice similarity coefficient(DSC)= where A is the manual contouring volume, and B is the auto-segmentation volume. DSC takes values between zero and one. When the DSC value approaches zero, the manual and auto-segmentation outcomes differ significantly. However, as the DSC value approaches unity, the two volumes exhibit increased similarities.   Dice similarity coefficient

14. the similarity of X and Y is determined according to the distance of the nearest maximum distance. HD is thus defined as, Hausdorff distance(HD)= max( d(X,Y),d(X,Y) Where d(X,Y) is the directed HD from X to Y and is given by d=max( min(||x-y||) xX , and yY As the HD approaches zeros, the difference between the manual contouring and auto contouring becomes smaller. By contrast, if the coefficient is greater than zero, the similarity between the two volumes decreases. Hausdorff distance(HD):

15. Mean distance to agreement(MDA) Mean distance to agreement is the mean voxel wise comparison of distance between two associative points in the contour sets A and B, defined by MDA(A,B) = mean a Є A,b Є B {d( a,B ) ud ( b,A )} and denotes a measure of average similarity between two contour sets. A higher MDA between two sets A and B indicates the existence of regions of dissimilarity between the two sets, where a MDA of zero indicates that the sets A and B are identical. Because MDA represents an average across all points in the sets A and B, MDA is less sensitive than HD to small pockets of high dissimilarity.

16. RVD compares the sizes between two volumes. RVD relative volume difference = As RVD approach zero, the manual contouring and auto contouring volumes only yield small volume differences, and values larger than zero reduce the similarity between the two volumes. A =manual contour , B= auto segmentation   Relative volume difference(RVD):

17. Statistical analysis Null hypothesis 1 : Geometrically, there is no significant difference in two systems i.e smartAdapt and velocity. Null hypothesis 2 : Dosimetrically, there is no significant difference between manual contours and auto-segmented contours. The Wilcoxon signed-rank test was used to compare the DSC,HD,MDA and RVD values of smart adapt and velocity. The Wilcoxon signed-rank test was used to compare between dosimetric parameters of auto segmented OAR’s with respect to the manual clinical contour. A p-value of less than 0.05 was considered statistically significant.

18. RESULTS & DISCUSSION Parameter (n=12) DSC MDA HD RVD SmartAdapt Velocity p value SmartAdapt Velocity p value SmartAdapt Velocity p value SmartAdapt Velocity p value Mean±SD Mean±SD Mean±SD Mean±SD Mean±SD Mean±SD Mean±SD Mean±SD Brainstem 0.81±0.05 0.77±0.05 0.239 1.76±1.76 1.99±1.99 0.388 8.66±3.48 8.44±1.98 1.000 0.23±0.17 0.16±0.12 0.248 Spinal Cord 0.81±0.04 0.74±0.07 0.008 0.97±0.97 2.17±2.17 0.008 6.23±3.52 16.54±18.72 0.041 0.18±0.16 0.14±0.13 0.410 Mandible 0.83±0.05 0.86±0.03 0.055 1.18±1.18 0.99±0.99 0.136 7.11±2.88 11.17±9.74 0.099 0.16±0.1 0.13±0.08 0.154 OralCavity 0.85±0.05 0.81±0.09 0.074 2.3±2.3 2.82±2.82 0.308 10.17±3.78 12.93±4.88 0.028 0.16±0.15 0.21±0.13 0.556 Parotid_R 0.73±0.06 0.7±0.05 0.155 2.05±2.05 2.17±2.17 0.433 12.73±2.66 14.41±5.9 0.530 0.28±0.21 0.16±0.12 0.155 Parotid_L 0.74±0.08 0.73±0.05 0.480 2.23±2.23 1.9±1.9 0.583 12.24±5.83 12.04±3.73 0.814 0.26±0.27 0.21±0.14 0.784 Eye_R 0.83±0.04 0.85±0.04 0.272 1.24±1.33 1.08±1.08 0.239 4.38±1.74 5.21±1.7 0.239 0.33±0.16 0.15±0.11 0.005 Eye_L 0.84±0.04 0.86±0.04 0.083 1.16±1.27 1.04±1.04 0.255 4.37±1.03 5.23±2.22 0.239 0.3±0.15 0.11±0.09 0.005 OpticNerve_R 0.39±0.22 0.36±0.1 0.638 1.72±1.72 1.51±1.51 0.937 7.39±2.87 6.38±2.72 0.328 0.45±0.29 0.34±0.19 0.433 OpticNerve_L 0.21±0.16 0.43±0.12 0.003 2.98±2.98 1.42±1.42 0.013 9.98±4.19 7.54±2.58 0.050 0.33±0.51 0.41±0.21 1.000 Esophagus 0.14±0.2 0.34±0.23 0.050 15.06±15.06 11.32±11.32 0.508 41.37±20.21 32.66±19.05 0.445 0.61±0.34 3.92±6.81 0.423

19. HD difference for spinal cord in SmartAdapt Manual contoured Auto segmented contoured Manual contoured Auto segmented contoured HD difference for spinal cord in velocity

20. The dose volume histograms (DVHs), were calculated for manual contours and auto contours. Determined whether the differences between the manual and auto-segmented contours were statistically significant. Dosimetric analysis

21. Parameter (n=12) Constraints Referance SmartAdapt Velocity p value (Ref Vs SmartAdapt) p value (Ref Vs velocity)     Mean ± SD Mean ± SD Mean ± SD Brainstem Dmax 23.36±13.6 27.17±12.48 26.18±13.84 0.012 0.092 D1cc 14.84±13.06 16.22±13.13 16.89±12.82 0.071 0.209 Spinal Cord Dmax 38.12±3.43 38.36±3.32 38.47±3.17 0.367 0.388 D1cc 34.06±4.05 34.24±3.59 33.83±3.58 0.754 0.480 Mandible Dmax 61.68±9.85 60.53±10.36 61.81±8.85 0.015 0.814 D1cc 53.66±10.43 51.03±9.48 52.54±9.9 0.003 0.071 OralCavity Dmax 60.81±7.47 61.08±8.98 61.04±7.88 0.583 0.875 D1cc 48.33±9.08 51.94±10.99 50.7±10.56 0.084 0.060 Dmean 28.57±3.71 29.49±4.13 29.05±4.03 0.017 0.272 Parotid_R Dmax 64.46±6.18 64.59±6.11 63.14±5.95 0.534 0.002 V26Gy(%) 55.43±15.03 61.35±10.55 45.18±16.36 0.060 0.028 Dmean 28.98±10.03 31.17±7.35 28.34±9.61 0.272 0.754 Parotid_L Dmax 66.83±6.06 67.2±6.11 65.96±5.83 0.237 0.015 V26Gy(%) 59.53±15.92 63.46±14.03 50.3±20.42 0.034 0.004 Dmean 31.61±11.1 33.34±8.91 30.62±10.67 0.239 0.327 Eye_R Dmax 2.79±2.06 2.42±1.97 2.92±3.02 0.373 0.433 D1cc 1.49±0.55 1.47±0.6 1.5±0.62 0.456 0.366 Dmean 1.18±0.35 1.11±0.33 1.15±0.37 0.007 0.782 Eye_L Dmax 2.68±2.02 3.03±2.18 3.27±2.74 0.875 0.505 D1cc 1.45±0.63 1.66±0.79 1.6±0.71 0.075 0.556 Dmean 1.14±0.37 1.15±0.37 1.19±0.39 0.077 0.929 OpticNerve_R     Dmax 1.57±0.56 1.51±0.55 1.52±0.48 0.119 0.581 D0.03cc 1.36±0.4 1.35±0.43 1.43±0.41 0.646 0.021 Dmean 1.29±0.39 1.28±0.37 1.25±0.33 0.431 0.195 OpticNerve_L     Dmax 1.61±0.64 1.37±0.5 1.62±0.68 0.003 0.906 D0.03cc 1.36±0.43 1.26±0.37 1.48±0.5 0.018 0.084 Dmean 1.29±0.43 1.24±0.42 1.24±0.36 0.285 0.272 Esophagus     Dmax 69.37±1.01 51.36±27.11 67.01±4.47 0.005 0.010 D0.03cc 68.18±0.9 50.64±26.94 66.22±5.02 0.007 0.541 Dmean 36.65±12.49 39.61±25.02 47.19±16.51 0.508 0.019

22. ▲ manual contours ■ auto-segmented contours Spinal cord mandible Oral cavity Parotid R Parotid L Brainstem DVH of SmartAdapt

23. ▲manual contours ■ auto-segmented contours Spinal cord mandible Oral cavity Parotid R Parotid L Brainstem DVH of velocity

24. Time analysis The time required to perform the auto contouring and manual contouring was measured. SmartAdapt – ranges from 2 min 20 seconds to 4 minute 23 seconds Velocity – ranges from 2 minutes 21 seconds to 2 min 50 seconds Manual – ranges from 16 to 20 min

25. Discussions Studies have reported that it is essential to have at least a similar head position between atlas and sample patients if not a perfect chin match, for an acceptable mandible contour. By restricting chin to sternum notch distance to 7-10 cm in our atlas sets we overcame this challenge. Accuracy of oral cavity automatic contours were highly influenced by mandible or head position. Hence oral cavity is giving good accuracy. Atlas in both smartadapt and velocity were first created using stringent criteria of age (30-65 years), then created and atlas set with sample size of 30 patients. No substantial difference found from 50 patient sample. By increasing the number of atlas size to create atlas sets, there was no effect in geometric parameters. The inconsistency in relatively small organs for HN subjects may be attributed to the specific atlas selection method where global intensity similarity is used as the matching metric and consequently , the contributions of relatively small local regions are discounted. E xample : optic nerves

26. Discussions The geometric discrepancy between AS and MS contours may have been caused by the inconsistencies in MS contours. For example large variations in the superior-inferior ranges of cord. Cervical oesophagus is relatively low intensity structures in a low contrast region and also subject to movement. Hence segmentation becomes difficult and thus had the least DSC score and high HD value among all structures. The dosimetric differences of organs with low geometric accuracies such as optic nerves are relatively small between AS and MS because they are located distant to the target and high dose region. If an organ is located in a high dose region with low dose gradient, its dosimetric metrics may have high absolute values, but minor variation related to geometric shape change. A larger pool of patient samples in future studies would be beneficial to characterize the dosimetry performance of each individual structure.

27. CONCLUSION At least a similar setup is an essential pre-requisite to generate an acceptable set of automatic contours in single atlas based systems. The dosimetry performance not only depends on geometric accuracy, but is also heavily impacted by spatial dose distribution and gradient. The geometric measures alone were not sufficient to predict the dosimetric impact of segmentation inaccuracies on RT treatment plans.

28. References Cao M, Stiehl B, Yu VY, Sheng K, Kishan AU, Chin RK, Yang Y, Ruan D. Analysis of geometric performance and dosimetric impact of using automatic contour segmentation for radiotherapy planning. Frontiers in oncology. 2020 Sep 23;10:1762. Hu Y, Byrne M, Archibald‐ Heeren B, Thompson K, Fong A, Knesl M, Teh A, Tiong E, Foster R, Melnyk P, Burr M. Implementing user‐defined atlas‐based auto‐segmentation for a large multi‐centre organisation: the Australian Experience. Journal of Medical Radiation Sciences. 2019 Dec;66(4):238-49. van der Veen J, Gulyban A, Willems S, Maes F, Nuyts S. Interobserver variability in organ at risk delineation in head and neck cancer. Radiation Oncology. 2021 Dec;16:1-1. Lorenzen EL, Kallehauge JF, Byskov CS, Dahlrot RH, Haslund CA, Guldberg TL, Lassen-Ramshad Y, Lukacova S, Muhic A, Witt Nyström P, Haldbo -Classen L. A national study on the inter-observer variability in the delineation of organs at risk in the brain. Acta Oncologica . 2021 Nov 2;60(11):1548-54.

29. ACKNOWLEDGEMENTS: Mr. M Anil Kumar Mr. Raghavendra Hajare Ms. KK Sreelakshmi Dr. Rohit Vadgoankar Dr. Kiriti Chiriki Dr. Chandrasekhar pusarla

30. Thank you