ESTRO 2021 Abstract Book

S1658

ESTRO 2021

PO-1946 An incremental learning-based model for organ at risk in cervical cancer multi-scene radiotherapy D. Qian 1 , Z. Zhang 2 , R. Zhou 2 , R. Wang 1 , S. Wu 3 , S. Wang 2 , Z. Yang 2 1 Perception Vision Medical Technology, Clinical Research Department, Guangzhou, China; 2 Xiangya Hospital, Department of Radiation Oncology, Changsha, China; 3 Sun Yat-Sen University, School of Computer Science and Engineering, Guangzhou, China Purpose or Objective There are multiple scenes of radiotherapy for cervical cancer, such as after-loading and external irradiation, which have huge differences in the image domain. Using an independent model delineates organ at risk(OARs) for each scene is inefficient and fails to use information between scenes. Therefore, an incremental learning- based model(ILBM) for OARs in multi-scene radiotherapy for cervical cancer is put forward. Materials and Methods ILBM adopted a three-branch structure, the after-loading branch pre-trained on the positive slices of after- loading Computed Tomography (CT) images with object organs of the rectum, bladder and sigmoid, the external irradiation branch, structure same as the after-loading branch, pre-trained on the positive slices of irradiation CT images with object organs of rectum and bladder, the fusion branch merging the after-loading irradiation branch as feature extract module with the main part, structure the same as an after-loading branch, trained on full data including after- loading and external irradiation CT images with full object organs above. During training, ILBM accepted after-loading and external irradiation data as inputs alternately. Dynamic weighted multi-task loss function among multiple organs and samples was proposed to guide the model to focus more on the difficult samples and organs. ILBM was trained on 80 cervical cancer patients with after-loading CT images, 41 with external irradiation CT images, evaluated on 4 after- loading and 5 external irradiation CT images. The Dice’s coefficient(DSC) was used to evaluate the accuracy of ILBM. Results On the testing dataset, ILBM achieved superior performance on most organs than Res-Unet under the same data augmentation strategy, with the mean DSC of 76.5(±6.18)% on the rectum, 89.6(±1.92)% on the bladder, 63.9(±1.3)% on the sigmoid of after-loading CT images, 47.2(±30.769)% on the rectum, 84.5(±6.47)% on the bladder of external irradiation CT images compared to Res-Unet with 75.7(±6.7)%, 88.1(±1.99)%, 60.5(±4.0)%, 40.7(±34.4)%, 85.6(±7.01)%. As illustrated in figure1, the autocontouring results by ILBM marched better with the gold target(GT) than results of Res-Unet.

Figure1: OARs segmentation comparison between ILBM(Second row) and Res-Unet (First row). Dark line:GT, Light line: prediction Conclusion The quantitative and visualization evaluation results demonstrated the effectiveness of ILBM, which adopted an incremental learning strategy with two branches pre-trained on after-loading and external irradiation positive slice CT images respectively to afford organ features on the respective scene and furtherly fused to ILBM to make better segmentation decision, alternation training strategy to guide model to learn the similar features in and between scene to push ILBM to converge quickly. PO-1947 catching opportunity from voxel evaluation of follow-up PET/CT imaging in SBRT of lung lesions. F. Cavallo 1 , L. Capone 2 , N. Gennuso 1 , G. Abate 3 , G. Grimaldi 4 , S.A. Allegretta 1 , I. Russo 1 , P. Gentile 2 1 UPMC HILLMAN CANCER CENTER VILLA MARIA, RADIOTHERAPY, MIRABELLA ECLANO, Italy; 2 UPMC HILLMAN CANCER CENTER SAN PIETRO FBF, RADIOTHERAPY, ROME, Italy; 3 UPMC HILLMAN CANCER CENTER SAN PIETRO FBF, RADIOTHERAPY, MIRABELLA ECLANO, Italy; 4 UPMC HILLMAN CANCER CENTER SAN PIETRO, RADIOTHERAPY, ROME, Italy Purpose or Objective In order to define an effective outcome in lung patients undergoing SBRT it is mandatory to evaluate post treatment imaging following solid standards. In recent times, the spreading of nuclear medicine and PET/CT, allow to use PET imaging for the follow-up assessment and its Standard Uptake Values (SUV) as biomarker. The aim of this study is to investigate the opportunity to use a voxel framework method from pre/post treatment SUVs in lung patients after SBRT for a rapid outcome evaluation. Materials and Methods In this preliminary analysis were involved five oligometastatic lung patients treated with SBRT protocols. Patients were selected with several prescribed doses (45 Gy – 60 Gy) and in a different time after radiotherapy (4-24 months). After Radiation Oncologists outcome evaluations following PERCIST criteria (Complete Metabolic Response-CMR; Partial Metabolic Response-PMR; Stable Metabolic Disease-SMD; Progressive Metabolic Disease-PMD), RTTs create a voxel map with pre/post SUV values, using Follow-up PET images. Regions of interest to determine 2D voxel maps were created from PTVs