Deep Learning for Chromosome Segmentation with Uncertainty Estimation

Sammanfattning: Karyotyping, the process of pairing and ordering chromosomes, is an important tool for cytogenetic analysis to detect chromosome abnormalities. A trained specialist analyses the resulting image of the karyotyping, known as a karyogram, paying attention to the size, shape, and number of the chromosomes. However, manual inspection is a time-consuming task and necessitates domain expertise. Thus, there is a great interest in automating this process to allow more individuals access to clinical genetics. One important piece of this puzzle is segmentation, whereby the individual chromosomes are delineated and separated from the background. Nevertheless, this becomes difficult when there are overlapping chromosomes in the original micrograph. Furthermore, most employed deep learning models are unable to alert the user when the segmentation is likely to fail. This thesis aims to bring clarity to the aforementioned issues. Firstly, an augmented dataset consisting of overlapping chromosomes is created, which is then used to train a convolutional neural network inspired by the original U-Net implementation, but with dropout layers, batch normalization and padding to ensure equal sizes of input and output. Our proposed model achieves a validation accuracy of 97.4 %. Different uncertainty metrics were then compared in terms of predictive capacity of the segmentation accuracy, both qualitatively from the uncertainty maps and quantitatively by computing the correlation. The results showed that the uncertainty obtained with Monte Carlo Dropout and Test Time Augmentation, both measured using entropy, were the most promising approaches to predict the accuracy. The uncertainty maps were used as training data on a regression problem with a ResNet network as our model to predict segmentation accuracy, where we could not demonstrate any significant benefit of the uncertainty estimations compared to the benchmark models.