Evaluation of transferability of Convolutional Neural Network pre-training with regard to image characteristics

Abstract: This study evaluates the impact of pre-training on a medical classification task and investigates what characteristics if images affect the transferability of learned features from the pre-training. Cardiotocography (CTG) is a combined electronic measurement of the fetal heart rate (FHR) and maternal uterine contractions during labor and delivery and is commonly analyzed to prevent hypoxia. The records of FHR-signals can be represented as images, where the time-frequency curves are transformed to color spectrums, known as spectrograms. The CTU-UHB database consists of 552 CTG-recordings, with 44 samples of hypoxic cases, rendering a small data set with a large imbalance between the hypoxic class versus normal class. Transfer learning can be applied to mitigate this problem if the pre-training is relevant. The convolutional neural network AlexNet has previously been trained on natural images with distinct motifs, including images of flowers, cars, and animals. The spectrograms of FHR-signals are on the other hand computer generated images (synthetic) with abstract motifs. These characteristics guided the selection of benchmark data sets, to study how beneficial the AlexNet pre-training is with regard to the characteristics. 5-fold cross-validation and t-tests with a 1% significance level were used for performance estimations. The ability to classify images from the benchmark data sets was significantly improved by pre-training, however not for the FHR-spectrograms. Varying the balance between classes or amount of data did not produce significant performance variations in any of the benchmark data sets, and each of them significantly outperformed the FHR-data set in all trials. Attempts to replicate previous results were unsuccessful. The suspected causes are methodological differences regarding preprocessing of the FHR- signals, differences in the AlexNet implementations and testing method. The performance when classifying the FHR-spectrograms was, therefore, unable to be validated. In conclusion, the results indicate that the AlexNet pre-training could generalize to synthetic images and improved performance for the benchmark data sets. Pre-training on natural images with distinct motifs does, however, not seem to contribute to an increase in model performance when classifying FHR- spectrograms. Pre-training and/or comparing with alternative spectrogram images is recommended for future research.