Aggregating predictions of a yeast semantic segmentation model : Reducing a pixel classifier into a binary image classifier

Abstract: The introduction of machine learning in clinical microbiology is important for aiding clinical laboratories with highly repetitive tasks that are fatiguing, error-prone, and require long employee training time due to the complex nature of the task. A challenging task that belongs to the subareas that need assistance is yeast detection in fluorescence microscopy where various yeast morphologies exist. Developing a model does not completely alleviate this problem since it does not necessarily mean that it will be used in practice. Interpretable machine learning studies have suggested that the acceptance of a machine learning product is related to the understanding of the product. Therefore, the aim of this master's thesis was to address the interpretability of a semantic segmentation model that was developed for yeast detection with a new approach. In this study, methods were developed for aggregating prediction masks into binary labels to determine if aggregated semantic segmentation models could operate the same function as a binary image classification model while allowing for higher interpretability. This was determined by comparing the differences in performance between the semantic segmentation model developed with transfer learning using a pre-trained VGG19 as the base model and a U-Net as the classification head against a benchmark model developed with a pre-trained VGG19. For each aggregation method, the McNemar test was performed to determine if there were any performance differences between each aggregation method and the benchmark model. The test revealed a statistically significant difference between them at significance level α = 0.01. Additionally, the benchmark model VGG19 achieved a perfect performance with 100% accuracy, recall and precision, while all aggregation methods achieved noticeably worse results. It was deduced that the poor results of the aggregation methods were likely connected to the VGG19-U-Net's strong bias toward negative segments. The results of this study were inconclusive: therefore, further research must be conducted. Areas for improvement in future research include better dataset collection focusing on equal-sized subsets of each yeast morphology to improve model performance, and rebuilding the model architecture with a multiclassifier head and a quantification ability. This will enable the development of a leaner and better-understood aggregation method.