ESTRO 2021 Abstract Book

S1384

ESTRO 2021

Purpose of this study was to perform an in-depth analysis of different TL approaches, with a focus on similarity between source and target datasets, in the automatic classification of chest radiographs in positive or negative COVID19 infection. The aim was to gain knowledge on TL using publicly available datasets for future AI applications in oncology. Materials and Methods To create the target dataset, 6000 digital chest X-rays were extracted from the publicly available BIMCV- COVID19 dataset, equally balanced between COVID19 positive and negative. As source datasets, ImageNet and a balanced binary dataset of 48.10 3 chest radiographs created from the ChestX-ray14 dataset were used. All datasets are illustrated in Figure 1. Of the target dataset, 10% was set aside as test set for all experiments. Pre-processing of all medical images consisted of normalization and adaptive histogram equalization. NNs VGG19 and ResNet50, as implemented in Keras (144.10 6 and 26.10 6 weights, respectively), were used to perform the experiments and evaluate possible network size dependence. Hyperparameters of both networks were optimized for the particular source and target task and for each TL approach separately. TL was applied with each source dataset and by fine-tuning all NN layers during target training. Classification performance on the target test set was compared to using no TL (random initiation), for different target dataset sizes ranging from 100 to 6000 images. The area under the receiver-operating curve (AUC) was used as performance metric for all experiments.

Results Figure 2 displays the AUC obtained on the target test set for the different TL approaches and using no TL, as a function of dataset size. For datasets smaller than 1000 images, no difference in classification performance can be observed between the different approaches. For datasets with over 1000 images, TL using ImageNet outperformed the other approaches with both NNs. Despite both NNs reaching high AUC on the ChestX-ray14 source test set (AUC test = 0.90), using a similar source dataset as the target task did not result in improved COVID19 classification.