ESTRO 2020 Abstract book

S874 ESTRO 2020

PO-1527 When small is too small? Training Deep Learning models in limited datasets. G. Valdes 1 , M. Romero 2 , Y. Interian 2 , T. Solberg 1 1 University of California UCSF, Radiation Oncology, San Francisco CA, USA ; 2 University of San Francisco, Data Science, San Francisco, USA Purpose or Objective To perform an in depth evaluation of current state of the art techniques in training neural networks to identify appropriate approaches in small datasets. Material and Methods 112,120 frontal-view x-ray images from the NIH ChestXray14 dataset were used in our analysis . Two tasks were studied: unbalanced multi-label classification of 14 diseases, and binary classification of pneumonia vs non- pneumonia . All datasets were randomly split into training, validation and testing (70%, 10%, 20%). A popular convolution neural network (CNN), DensNet121, was trained using PyTorch. We performed several experiments to test: 1) whether transfer learning using pre-trained networks on ImageNet are of value to medical imaging / physics tasks (e.g., predicting toxicity from radiographic images after training on images from the internet), 2) whether using pre-trained networks trained on problems that are similar to the target task helps transfer learning (e.g., using x-ray pre-trained networks for x-ray target tasks), 3) whether freezing deep layers or changing all weights provides an optimal transfer learning strategy, 4) the best strategy for the learning rate policy and, 5) what quantity of data is needed in order to appropriately deploy these various strategies (N=50 to N=77,880). Finally, linear models using radiomics features (79 in total) were trained as baseline and confident intervals were calculated using bootstrap. Results In the multi-label problem, DensNet121 needed at least 1600 patients to be comparable to, and 10,000 to outperform, radiomics-based logistic regression. In classifying pneumonia vs. non-pneumonia, both CNN and radiomics-based methods performed poorly when N < 2000. For small datasets (< 2000), however, a significant boost in performance (> 15% increase on AUC) comes from a good selection of the transfer learning dataset, dropout, cycling learning rate and freezing and unfreezing of deep layers as training progresses. In contrast, if sufficient data is available (>35000), little or no tweaking is needed to obtain impressive performance. While transfer learning using x-ray images from other anatomical sites improves performance, we also observed a similar boost by using pre-trained networks from ImageNet, Figure 1. Having source images from the same anatomical site, however, outperforms every other methodology, by up to 15%., Figure 1 In this case, DL models can be trained with as little as N=50, Figure 1.

The median OS calculated from diagnosis was 21 months for patients without metastases and 13 months for patients with metastases. Higher LDH (p<0.001) and NSE (p<0.001) correlated with worse OS. The expression of synaptophysin correlated with a better OS (HR 0.529 95% CI 0.299-0.937, p=0.029). The expression of TTF-1(HR 0.282, 95% CI: 0.116- 0.684, p=0.005) and a lower GLUT-1 H-score (median = 50, HR: 0.524, 95% CI: 0.273-1.003, p=0.05) correlated with a better PFS. Chromogranin A correlated with the presence of cerebral metastases (OR 0.268, 95% CI: 0.087-0.826, p=0.02). Our radiomics analysis did not reveal a single texture feature that was highly correlated with overall survival or progression free survival. Correlation coefficients ranged between -0.35 and 0.45 for overall survival and between -0.27 and 0.49 for progression free survival. Conclusion In this analysis the expression of synaptophysin correlated with a better OS. This marker might be considered for decision making such as treatment intensification in synaptophysin-negative patients. No radiomic features correlated with any of the markers or the outcome of the treatment. PO-1526 a radiomics signature to predict response of chemoradiotherapy in esophagus squamous cell carcinoma B. Li 1 , Q. Cao 2 , D. An 3 1 Shandong Cancer Hospital and Institute, Department of radiation oncology, Jinan, China ; 2 Southeast University, Laboratory of Image Science and Technology, Nanjing, China ; 3 Shandong University, Cheeloo college of medicine, JInan, China Purpose or Objective To investigate the potential image markers for early prediction of treatment response on thoracic esophagus squamous cell carcinoma (ESCC) treated with concurrent chemoradiotherapy (CCRT). Material and Methods 159 thoracic ESCC patients enrolled from two institutions were divided into training and validation sets. A total of 944 radiomics features were extracted from pretreatment 18 F-FDG PET image sequences. We first performed the inter-observer test (two delineate methods) and limma on 10 pairs of patients (responders vs. nonresponders) to identify repeatably differentially expressed features (DEFs). Then, the least absolute shrinkage and selection operator (LASSO) logistic regression model with 10-fold cross-validation was used to construct a treatment response related radiomics signature. At last, the performance of this signature was assessed in both sets with receiver operating characteristic (ROC) curves and Kaplan-Meier (KM) analysis. Results After the inter-observer test, 691 features under two delineate methods were left ( p <0.05, ICC >0.9). 61 of them from limma were regarded as DEFs and entered the LASSO model. The 7-feature radiomics signature was significantly associated with treatment response ( p <0.001 in the training set and p =0.026 in the validation set) and achieved an area under curve (AUC) value of 0.844 and 0.835 respectively. Delong test of two ROCs showed no significant difference ( p =0.918). The cut-off value of radiomics signature could successfully divide patients into high-risk and low-risk groups in both sets. Conclusion That study indicated that the proposed radiomics signature could be a useful image marker to predict the therapeutic response of thoracic ESCC patients treated with CCRT.

Conclusion While training DL models in small datasets (N<2000) is challenging, no tweaking is necessary for bigger datasets