Oral Presentation 37th TROG Cancer Research Annual Scientific Meeting 2025

Development of an automated contour quality assurance tool for the TROG 08.08 TOPGEAR trial  (#47)

Phillip Chlap 1 2 3 , Mark Lee 3 , Trevor Leong 4 , Matthew Field 1 2 3 , Jason Dowling 2 5 , Hang Min 1 2 5 , Julie Chu 4 , Jennifer Tan 4 , Phillip Tran 4 , Tomas Kron 4 , Annette Haworth 6 , Laurence Court 7 , Martin A Ebert 8 9 10 , Shalini K Vinod 1 2 3 , Lois Holloway 1 2 3 6
  1. Ingham Institute of Applied Medical Research, Liverpool, NSW, Australia
  2. University of New South Wales, Sydney, NSW, Australia
  3. Liverpool and Macarthur Cancer Therapy Centre, Liverpool, NSW, Australia
  4. Peter MacCallum Cancer Centre, Melbourne, VIC, Australia
  5. Australian e-Health Research Centre, CSIRO, Brisbane, QLD, Australia
  6. Institute for Medical Physics, School of Physics, The University of Sydney, Sydney, NSW, Australia
  7. Department of Radiation Physics, University of Texas MD Anderson Cancer Center, Houston, TX, USA
  8. Sir Charles Gairdner Hospital, Perth, WA, Australia
  9. University of Western Australia, Perth, WA, Australia
  10. University of Wisconsin, Madison, WI, USA

Background
The TROG 08.08 TOPGEAR trial investigated the use of preoperative chemoradiotherapy in addition to perioperative chemotherapy for patients with resectable gastric cancer. The TOPGEAR protocol defined a Clinical Target Volume (CTV) to receive the prescribed radiation dose. Manual contour quality assurance (CQA) was performed on a subset of patients throughout the trial to ensure conformance of the CTV with the trial protocol.

Aims
In this work we aim to develop an approach for automating CQA for the CTV structure within the TOPGEAR trial.

Methods
The CTV structure was defined by 5 radiation oncologists on 10 cases. To account for contour interobserver variation uncertainty, auto-segmentation models were trained for both the intersection and the union of the 5 observers. These models were then applied to a set of 40 cases treated within the TOPGEAR trial. A metric was defined  by the proportion of the manual CTV contour points that fall within the predicted intersection and union auto-segmentations. Although these 40 cases had undergone review during the trial, an additional review was performed to determine the CQA result at the end of the trial.  Each case had a passing and/or failing CTV available. These results were compared to the output of the automated approach by thresholding the metric defined at various levels to maximize an Area Under the Receiver Operating Characteristic Curve (AUC-ROC).

Results
An AUC-ROC score of 0.81 was achieved. When selecting the threshold for which 31/33 failing CTVs are correctly flagged, 15/32 passing CTVs are also flagged. This model results in a sensitivity of 0.94 and specificity of 0.53.

Conclusions
An approach was developed to automatically flag CTVs not conforming to the TOPGEAR trial protocol. Future work will address uncertainty within the model to help determine for which cases the automated approach is reliable.