ESTRO 2021 Abstract Book

S1530

ESTRO 2021

to controls. The LET d was also consistently higher for all the structures. While the confidence intervals overlapped, the differences between cases and controls in the patient cohort could be a contributor to the brainstem toxicity, however, with small effects compared to other patient specific factors or interpatient radiosensitivity variability.

PO-1802 Deep learning using Pre-NACRT imaging can predict pathological response in esophageal cancer K.S. Chufal 1 , I. Ahmad 1 , A. Dwivedi 2 , R. Bajpai 3 , A.A. Miller 4 , R.L. Chowdhary 1 , M. Gairola 1 1 Rajiv Gandhi Cancer Institue & Research Centre, Radiation Oncology, New Delhi, India; 2 Almini Services Limited, Software Architecture, Reading, United Kingdom; 3 Keele University, School of Medicine, Staffordshire, United Kingdom; 4 Illawarra Cancer Care Centre, Radiation Oncology, Wollongong, Australia Purpose or Objective To develop a deep learning Convolutional Neural Network (CNN) to predict complete pathological response (pCR) after Neo- Adjuvant ChemoRadioTherapy (NACRT) in oesophageal cancers, based on pre-NACRT imaging alone. Materials and Methods From our institutional dataset of oesophagal cancer patients who underwent NACRT followed by surgery (N=254), 211 patients had pre-NACRT DICOM CT imaging data available. 192 sequential patients were utilised for model development (training & validation sets contained 154 and 38 patients, respectively) and validated on 19 patients. Deep Learning Methodology Data pre-processing & Augmentation All tumours were delineated on Pre-NACRT CT imaging and converted to NIFTI-2 format. After resampling into isotropic voxels, the delineated 3D tumour's centre-of-mass served as the origin for a 64x64x64 bounding box, upon which an augmentation factor of 6000 was applied. Model Building The base 2D model architecture of Visual Geometric Group (VGG-16) was customised to handle 3D NIFTI-2 data. Hyperparameter (HP) optimisation Optimisation was performed in three steps: 1. Random Search: 50 randomly generated sets of HPs were set on 3 optimisation runs per set to establish the prior probability distribution. 2. Bayesian Optimisation: Based on the prior probability distribution, a Bayesian Optimiser (BO) maximised an objective function (QMetric). The BO was set on 100 runs of 100 epochs each, and the HPs of 10 best models were used to create a range for each HP, upon which a grid search was performed. 3. Grid Search: The range for each HP was exhaustively searched using all possible combinations. Statistical Analysis The final model's performance was evaluated using augmented and non-augmented training and validation sets with pCR as the outcome. Results CNN Architecture The network comprised nine 3D convolutional layers in four groups. A leaky ReLUs activation function followed the first three groups of layers, and a max-pooling layer was applied after the last group. After flattening, three fully connected dense layers culminated in an output layer (a sigmoid function to predict pCR and pPR probability). We utilised the gradient-based stochastic optimiser Adam and a Cyclical Learning Rate.

Model Performance On evaluating the model's accuracy using training generator (25 runs), the mean accuracy was 0.75 ± 0.02 (Mean Loss: 1.18 ± 0.006; AUC = 0.724). The accuracy in the validation set was 0.74 (Mean Loss: 1.15; AUC = 0.697). On internal validation, the F1-score was 0.8 for predicting pCR and 0.25 for predicting pPR. The overall accuracy of the model was 0.68. We mapped the network’s activation maps over the final convolutional layer to visualise the regions deemed most relevant for predictions. The region within the GTV, excluding the oesophagus lumen, was most predictive of pCR.