Developing TomoNet: A Deep Learning Feature Extractor for Medical Image Classification

Abstract

In this thesis, I investigated whether incorporating medical images in the training of feature extractors would enhance their performance in medical imaging tasks. To achieve this objective, I conducted two studies. In the first study, I compared the performance of a feature extractor pretrained with natural images from ImageNet (referred to as ImageNet-derived) with a feature extractor trained from scratch without any pre-existing feature extraction capabilities. Both feature extractors were trained to classify medical images into nineteen categories. The results demonstrated that the performance of the feature extractor trained from scratch approached the performance of the ImageNet-derived feature extractor as the size of the training set increased. Additionally, I introduced a metric called the "transfer learning gap" to quantify the relative amount of knowledge transfer in transfer learning. The experiments revealed that training the feature extractor twice, first with the ImageNet dataset and then with 160,000 medical images, resulted in a better-performing deep learning model. This was an unexpected result and likely due to training with an insufficient number of medical images. Thus, I named the better performing, twice-trained feature extractor "TomoNet" and used it in my subsequent studies. In the second study, I compared the performance of TomoNet with that of the ImageNet-derived feature extractor in classifying medical images based on sex. This study failed to provide conclusive evidence that TomoNet, the twice-trained feature extractor, outperformed the ImageNet-derived feature extractor. Both models exhibited signs of overfitting, and there was potential evidence of catastrophic interference in TomoNet results. Finally, I presented an initial business model for a company aiming to commercialize products derived from the TomoNet studies. The business model focused on the commercialization of "Petal-Blue," a deep learning model designed to detect white matter hyperintensities associated with dementia in brain imaging. The Canvas Business Model framework was utilized to evaluate crucial aspects of the startup, including value proposition, customer segments, revenue streams, and partnerships.