ESTRO 2022 - Abstract Book

S119

Abstract book

ESTRO 2022

MO-0144 Deep learning to detect nuclei and DNA damage foci for ex vivo tissue radiation sensitivity analysis

I. Lauwers 1 , K. Pachler 2 , M. Capala 1 , N. Sijtsema 1,3 , M. Hoogeman 1,4 , G. Verduijn 1 , S. Petit 1

1 Erasmus Medical Center, Department of Radiotherapy, Rotterdam, The Netherlands; 2 Erasmus Medical Center, Department of Molecular Genetics, Rotterdam, The Netherlands; 3 Erasmus Medical Center, Department of Radiology and Nuclear Medicine, Rotterdam, The Netherlands; 4 HollandPTC, Department of Medical Physics and Informatics, Delft, The Netherlands Purpose or Objective Proton therapy may be more effective at killing tumor cells than photon therapy due to enhanced DNA damage. However, the underlying mechanisms are not well understood and may vary between patients. Therefore, we keep rest material from oral cavity tumors viable and study its responds to proton and photon irradiation ex vivo . The rest material is stained for DNA damage and imaged. However, manual counting of nuclei and thousands of DNA damage foci in the images is time consuming, preventing large scale quantitative comparisons. Accordingly, the purpose of this study was to develop an automated method to segment and count nuclei and DNA damage foci in tissue, to study the dynamics of DNA damage response after photon and proton therapy. Materials and Methods 319 images of tissue stained with DAPI (nuclei) and 53BP1 (DNA damage foci) were used. These images were obtained from resection rest material from seven tumors of patients operated for oral cavity squamous cell cancer (IRB MEC-2017-1049). Tumor tissue was sliced, cultured, treated with 0 or 5 Gy, stained at different time points and imaged under a Leica Stellaris 5 confocal microscope. The image size was 145.14 x 145.14 µ m 2 . To identify nuclei and foci, two separate convolution networks (U-Net) were trained (N=255) and validated (N=64). For training and validation purposes, masks of correct segmentations were created. The U-Nets were characterized by four layers with 64, 128, 256 and 512 filters, max pulling and normalization after each layer; 200 epochs for training; a batch size of 32 images; Jarcard index loss function and Adam optimizer (learning rate = 10 -3 ). For validation, the Dice index was used on a pixel by pixel basis for the nuclei images. Before counting the nuclei, hole filling, watershed (0.8), and size exclusion were performed. For 20 images selected at random, the number of foci and nuclei as determined by the model were compared to manual counting, as gold standard. The correlation was assessed using the coefficient of determination (R 2 ). Results The U-Nets were successfully trained (nuclei: Dice index=0.897) and post processed (Figure 1). One of the 20 nuclei images was excluded as outlier since it greatly differed from the other images. Figure 2 shows the automatically counted number of foci and nuclei vs the gold standard. An excellent correlation was observed between automated and manual counting for both nuclei (R 2 =0.97) and foci (R 2 = 0.93). On average the number of counted nuclei and foci deviated by less than 10% from the gold standard (automated/manual = 1.02 ± 0.07 (nuclei) and 1.09 ± 0.47 (foci).

Conclusion An accurate U-Net model was developed to automatically segment and count nuclei and DNA damage foci in tumor tissue. The high R 2 values and accurate average prediction demonstrate the excellent performance of the model. This model can