ADSN Inaugural Conference

Title: Analysing images using Deep Convolutional Neural Networks to supplement human decision making

Abstract: Recognising and classifying objects within images is generally a straightforward task for Deep Convolutional Neural Networks (DCNNs), with their performance exceeding humans in benchmark competition datasets since 2015. Their impressive performance on object images allow for relatively simple integration into automated systems. An open question is how well do these networks deal with more challenging datasets, such as those containing medical images, which would require a level of visual expertise in human observers? I describe a case study in collaboration with the UZLeuven hospital in Belgium, where we analysed a dataset of over 30,000 microscopy images of fertilised human embryos. To date, DCNNS have assisted in classifying embryos as early as day 5 after insemination. We investigated whether DCNNs could successfully predict the destiny of each embryo (discard or transfer) at an even earlier stage, namely at day 3. We first assessed whether the destiny of each embryo could be derived from technician scores, to examine whether the ratings that technicians gave to images correlated with the decision made. We then explored whether a DCNN could make accurate predictions using images alone. We found that a simple 8-layer network was able to achieve 75.24% accuracy of destiny prediction, outperforming deeper, state-of-the-art models. Importantly, when analysing cases of transferred embryos, we found that our lean, DCNN predictions were correlated (0.65) with clinical outcomes. I describe some of the known shortcomings of DCNNs compared to human observers, which is especially relevant when integrating this technology within a clinical context for making medical decisions.