ESTRO 2021 Abstract Book

S69

ESTRO 2021

Materials and Methods 163 subjects across 9 institutions were combined into one cohort for this evaluation. Ten specific structures were evaluated in this work: parotid (L and R), mandible, brainstem, eye (L and R), optic chiasm, optic nerve (L and R), and spinal cord. The structures were first auto-segmented using Contour ProtégéAI TM (MIM Software Inc.) (CPAI). Each test image was then evaluated and either edited or re-segmented by an expert clinician. These expert segmentations were compared against the original CPAI output to obtain the Dice Similarity Coefficient (DSC), mean Hausdorff Distance (HDmean), and Max Hausdorff Distance (HDmax). Results The average results across the whole dataset were as follows (DSC, HDmean, HDmax): Right Parotid (0.83, 1.69, 10.85), Left Parotid (0.83, 1.72, 11.07), Mandible (0.89, 0.81, 11.48), Brainstem (0.82, 1.67, 9.28), Right Eye (0.84, 1.72, 4.23), Left Eye (0.84, 1.66, 4.15), Optic Chiasm (0.63, 1.37, 4.8), Right Optic Nerve (0.77, 1.42, 7.04), Left Optic Nerve (0.74 0.8, 6.27), and Spinal Cord (0.88, 4.27, 27.12). 16% of contours in this dataset (144 out of 892 ) achieved a DSC of 0.99 or 1.0.

Conclusion Except for the optic chiasm and the optic nerves, all evaluated structures achieved at least 0.80 DSC, and the mandible achieved the highest average DSC (0.89). While the optic chiasm had the lowest average DSC (0.63), it is worth noting that 21% of optic chiasm contours in this cohort had a DSC of 0.99 or 1.0. It was inferred that for any contours with 0.99 or 1.0 DSC, the clinician used the CPAI output with no edits, meaning 16% of CPAI contours evaluated in this study needed no edits before clinical use. HD averages all fell within acceptable ranges, except for the spinal cord, which can be attributed to differences in superior/inferior contour extent. These results demonstrate overall that CPAI produces segmentations which are clinically usable in real-world scenarios. Future studies will repeat this analysis with more data, as well as evaluate trends and variability between institutions. OC-0095 ADC predicts persistent cervical lymph node disease following curative (chemo)radiotherapy A. Salah 1 , Y. Jain 2 , S. Bonington 2 , A. France 3 , D. Buckley 4 , C. Eccles 5 , A. McPartlin 6 1 The Christie NHS Foundation Trust, Proton Beam Therapy, Manchester, United Kingdom; 2 The Christie NHS Foundation Trust, Radiology, Manchester, United Kingdom; 3 The Christie NHS Foundation Trust, Proton Clinical Outcomes Unit, Manchester, United Kingdom; 4 The Christie NHS Foundation Trust, Medical Physics & Engineering, Manchester, United Kingdom; 5 The Christie NHS Foundation Trust, Radiotherapy, Manchester, United Kingdom; 6 The Christie NHS Foundation Trust , Head and Neck Clinical Oncology, Manchester , United Kingdom Purpose or Objective The value of diffusion weighted (DW)-MRI post (chemo)radiotherapy ((C)RT) to assess disease response in malignant cervical lymph nodes (LN) is incompletely understood. We assess the correlation between apparent diffusion coefficient (ADC) of cervical LN post (C)RT and subsequent neck dissection histopathology. Materials and Methods Patients treated with (C)RT for head and neck squamous cell carcinoma at a single institution with post- treatment MRI followed by surgical neck dissection from January 2017 - January 2020 were identified for this retrospective study. MRI scans were acquired at 1.5 T or 3 T on one of three scanners (Siemens Aera, Siemens Skyra, and GE Signa), and included structural sequences and DW-MRI with automatically generated ADC maps. The DW-MRI sequence parameters were comparable on all three scanners (b-values = 50 and 800 s/mm ² ). However, due to the limited field of view of the sequence, and the prioritisation of imaging primary disease, some of the DW-MRI scans did not include the full extent of neck nodes. Two experienced readers drew regions of interest (ROI) on the ADC maps over all identifiable cervical LN with a diameter greater than 5 mm, blinded to histopathological diagnoses, using the picture archiving and communication system (PACS). Mean and minimum ADC values (ADC mean and ADC min ) of each node were measured on the PACS workstation (by assessing the ADC of all pixels within the ROI) and matched to histopathological findings following neck dissection. An un-paired t-test was used to compare ADC values for benign and malignant cervical lymph nodes, with P < 0.05 considered statistically significant. Threshold ADC values for the detection of cervical LN malignancies were calculated using receiver operating characteristic (ROC) analysis. Results 31 cervical LN were identified in 18 patients. However, the ADC min for two of the LN was 0 mm ² /s, which may imply a failure in the ADC estimation, hence those two LN were excluded from analysis. For the remaining 29 nodes (14 malignant and 15 benign as confirmed on histology), there was no statistically significant difference between ADC mean for malignant (1.17 ± 0.36 x 10 ¯³ mm ² /s) vs. benign (1.40 ± 0.53 x 10 ¯³ mm ² /s) nodes (p =