ESTRO 38 Abstract book

S164 ESTRO 38

thousand-three-hundred-ninety rectangle regions were selected and annotated by pathologist with four categories including normal tissue, low, middle and high differentiation tumor. A convolutional neural networks (CNN) model which is similar to Xception model structure was used to training. Briefly, 256 * 256 pixels’ small patches with 40x magnification were extracted from WSI and assigned to their category. In model training, 152(90%) WSI images were used for training and 17 (10%) WSI images were used for validation. The initial learning rate is 1e-4 and the optimization algorithm is Adam. Each epoch has 1000 batches; each batch have 32 patches. The learning rate will decrease 10 times after every 150 epochs. An additional test WSI, which was not enrolled in model training and validation, was selected for demonstration our model’s results.

Radiation Oncology, Vienna, Austria ; 3 Institute Medic Onco Radioterapia, Department of Radiation Oncology, Barcelona, Spain ; 4 Antoine Lacassagne Cancer Center & University of Cote d’Azur, Biostatistics’ Unite, Nice, France ; 5 National Institute of Oncology, Department of Radiation Oncology, Budapest, Hungary ; 6 Catalan Institute of Oncology, Department of Radiation Oncology, Barcelona, Spain ; 7 Offenbach University Hospital, Department of Radiation Oncology, Offenbach, Germany ; 8 University of Kiel- Kiel, Department of Radiation Oncology, Kiel, Germany ; 9 University Hospital Berne- Berne- Switzerland, Department of Radiation Oncology, Berne, Switzerland; 10 University of Wurzburg, Department of Radiation Oncology, Wurzburg, Germany ; 11 University of Lubeck, Department of Radiation Oncology, Lubeck, Germany; 12 Experimental Radiotherapy Laboratory, Department of Radiation Oncology, Leuven, Belgium ; 13 University Erlangen- Nuremberg, Department of Radiation Oncology, Erlangen, Germany Purpose or Objective In case of second ipsilateral breast tumor event (2 nd IBTE) occurring after primary breast conserving surgery, salvage mastectomy or 2 nd conservative treatment (2 nd CT) including breast conserving salvage-surgery and salvage accelerated partial breast re-irradiation (APBrI) with brachytherapy can be performed. We report update results of 2 nd CT from the database of the GEC-ESTRO Breast Cancer Working Group. Material and Methods Between 2000 and 2014, 331 patients (pts) underwent a 2 nd CT in 12 hospital/cancer centers from 7 European countries. After salvage-lumpectomy, APBrI was performed using either low (30 – 55 Gy reference isodose) or high dose-rate brachytherapy (28 - 34 Gy). Oncological outcome including 3 rd IBTE, regional (RFS) or metastasis- free survival (MFS), specific (SS) and overall survival (OS) was analyzed. The belonging to a specific group of the GEC-ESTRO APBI classification (GAC) was also investigated. Simultaneously late side effectsand prognostic factors for 3 rd IBTE were analyzed. Results With a median follow-up of 72 months (range: 67 - 80 months), 143 pts (43%), 140 pts (42 %) and 48 pts (15%) were classified as low (LR), intermediate (IR) and high-risk (HR) respectively. For the whole cohort, 6-year 3 rd IBTE free survival, RFS, MFS, SS and OS rates were 92.9%, 96.4%, 87.4%, 90.1% and 85.8% respectively.6-year 3 rd IBTE free- survival rates for LR, IR and HI were 99.3%, 90.4% and 92% respectively (p = 0,009). 6-year RFS, MFS, SS and OS rates according to GAC are reported in Table 1. In UVA, SBR (1,2 vs 3 - p = 0.008), age (< 50 y vs.³50 y - p = 0.002) and GAC (p = 0.009) were considered as significant prognostic factors for 3 rd IBTE, while, in MVA, SBR (p = 0.046) and GAC (p = 0.01) were the two remaining prognostic factors. In terms of late toxicity, 194 pts (87%) presented G1,2 complications while G³3 complication rate was 13%.

Fig1. Workflow of our current works.

Results The results were presented in table 1. The category accuracy of predicting small patch was 81.6% for training set and 79.6% for validation set. The model has consumed 8 days to research this accuracy. Figure 2 demonstrated a results of test WSI. Table.1 Performance (accuracy) of model for series classes.

Figure 2. The pathology images and category heatmaps. a. Origin pathology image; b. Low differentiation; c. Medium differentiation; d. High differentiation; e. Normal tissue. f. Result of classification Conclusion Our model can precisely distinguish tumor and normal tissue on small patches. The accuracy for single WSI may further improve by combining these patches results. And deep-learning model can assist pathologists in the detection cancer differentiation. Meanwhile it can be used for prognosis prediction and assist decision making in the future study.

3 rd IBTE-FS RFS MFS SS OS

6 years

Whole cohort 92.9

96.4 87.4 90.1 85.8 94.4 87.2 92.0 87.8 92.7 89.6 91.8 86.7

APBI LR APBI IR APBI HR

99.3 90.4 92.0

Proffered Papers: BT 4: Breast and Skin brachytherapy

88.9 81.0 80.4 77.8 Table 1: Oncological outcomes for the whole cohort and according to the GEC-ESTRO APBI classification Conclusion In case of 2 nd IBTE, 2 nd CT combining re-lumpectomy + APBrI represent a valid therapeutic option in terms of oncological outcome as well as toxicity profile. Patient and tumor characteristics have to be carefully evaluated

OC-0317 2nd Conservative Treatment for 2nd Breast Tumor Event: GEC-ESTRO Breast Cancer WG updated results J. Hannoun-Levi 1 , K. Daniele 2 , G. Benjamin 3 , G. Jocelyn 4 , S. Renaud 4 , P. Csaba 5 , G. Cristina 6 , N. Pieter 7 , G. Ravzan 8 , L. Kristina 9 , P. Bulent 10 , K. Georgy 11 , V.L. Erick 12 , S. Vratislav 13 1 Centre Antoine Lacassagne, Radiation Therapy, Nice, France ; 2 Medical University of Vienna, Department of