A Comparative Study on the effect of different hyperparameters on the performance of VGGNet-16 for detection of Cardiomegaly in Chest X-ray Images

Sammanfattning: Computer aided diagnostics (CAD) systems have been widely researched and used in the medical field since it was introduced in the 1960s. The system functions as a support for radiologists in medical examinations using imaging technology such as X-rays, MRI and CT scans to diagnose diseases and treat injuries. As technology has developed throughout the years concepts such as neural networks have emerged and increased in popularity in this field. In the domain of neural networks, deep learning methods such as convolutional neural networks (CNNs) have shown promising and impressive results in many areas including image recognition. Increased development has led to using CNNs in CAD systems to function as an improvement to existing systems. A well performing CAD system using a CNN relies on a well performing CNN model. Obtaining a high performing CNN model is no easy task since it heavily depends on finding the right hyperparameters and a sufficient dataset. Several approaches on finding well suited hyperparameters exist today which includes hyperparameter tuning algorithms and manual trial and error. The goal for this thesis was to evaluate the effect of different configurations of hyperparameters on a specific CNN model’s performance. A promising and popular CNN model called VGGNet-16 was used in the study to diagnose a health condition called Cardiomegaly where a patient suffers from an enlarged heart using chest X-ray images. The hyperparameters chosen were the learning rate, batch size and optimizer. A dataset of X-ray images labeled with "Cardiomegaly" respectively "No findings" was used for training the model with different values of the hyperparameters for each training session. The results showed that the learning rate had the greatest impact on the model’s performance and therefore can be seen as the most important hyperparameter. The choice of optimizer showed no drastic effect on the model’s performance and achieved a similar accuracy rate given the right values on batch size and learning rate. Further it was found that a lower value of the learning rate generally gave higher performance compared to higher values and the combination of a low learning rate and low batch size is preferable for getting an increased performance of the model.