Automated acquisition of microscopic images of cells enables capturing their changes resulting from various conditions. These captured changes are quantitatively represented by a wide range of features called profiles, which can be obtained in different ways. In this master’s thesis we used profiles derived through deep learning, which, compared to hand-crafted features, are non-intuitive, and focused on improving their interpretability. We compared manual features with deep features extracted from the early, intermediate, and late layers of the deep model architecture. Through an ablation study, performed by averaging individual channels of the input images, we investigated how changes in the input data affect the resulting profiles. This design shows that deep features are better at detecting subtle differences between treatments than hand-crafted features, with intermediate layers most reliably capturing morphological differences of cells, while later layers focus more on classification and therefore explain general patterns less effectively. The ablation analysis reveals that the most informative channels are those of the endoplasmic reticulum (ER) and mitochondria (Mito), as averaging them causes the largest drop in accuracy across classification and analytical tasks.
|
