Mutations of the DES gene can occur de novo but can also be inherited. These mutations cause the desmin protein to agglomerate within muscle cells instead of fulfilling its important role of providing structural integrity for muscle cells during contraction. Desmin also acts as a scaffold for the cell architecture, i.e. the locations of the functional organelles such as the nuclei. Affected individuals typically suffer from muscle weakness, but can also experience cardiac arrest or respiratory failure, which can lead to death. The state-of-the-art method for analyzing myopathies in research and also diagnosing biopsies in clinical environments includes manual labeling of muscle cells in images of muscle samples, which is a time-consuming process. The extracted parameters can be used to study the efficacy of treatments in research or to support diagnosis in clinics. The time-consuming process of manual labeling and analysis slows treatment development. Here, we present a workflow that decreases the time required to analyze samples by fully automatizing the labeling process and parameter extraction steps. Further, this allowed us to analyze a greater amount of muscle cells than manual labeling would have allowed in a reasonable time frame. The results obtained allowed us to gain novel insights into the desmin knockout mouse model, a mouse model used to research desminopathies in humans. First, we trained a neural network with the U-Net architecture, U-Net-Real, using a supervised learning approach to automatically segment muscle cells in region of interest (ROI) images of Hematoxylin and eosin (H&E) stained formalin-fixated paraffin-embedded (FFPE) muscle samples (dataset A). The performance of U-Net-Real was evaluated using the gold standard evaluation metrics such as the accuracy (0.93). We then implemented a workflow to post-process and analyze the predictions obtained from U-Net-Real as a Fiji macro. Due to the satisfactory results, we decided to apply the approach to a different dataset, dataset B, consisting of whole slide images of H&E stained cryosectioned posterior compartment of the leg (calf muscle) samples. However, we found that U-Net-Real did not perform well on dataset B due to a domain shift in the datasets. Using synthetic images generated through simulations of real images adjusted for the domain shift via the novel SYNTA approach Mill 22, we trained a U-Net, U-Net-Synth, and applied it to dataset A and dataset B again. In comparison to U-Net-Real and another segmentation method, Cellpose (accuracy = 0.93), segmentation by U-Net-Synth on dataset A showed only a minimal performance loss (accuracy = 0.92). Visual inspection of the segmentation performance of U-Net-Synth and U-Net-Real on images of dataset B revealed better performance of U-Net-Synth. We then adjusted the analysis pipeline within the Fiji macro to account for new artifacts and tissue types that were visible due to the whole slide imaging approach used for dataset B. Further, we implemented a novel connective tissue mask generation algorithm, that allowed us to analyze the connective tissue in addition to the muscle cells. To assess the performance of U-Net-Synth segmentation, the Fiji macro analysis pipeline, and connective tissue mask generation, we carried out an expert evaluation. All three aspects combined had an average rating of 4.07 out of 5 points confirming the performance of the entire workflow. Hence, we applied the workflow to all images from dataset B. In total, dataset B consists of 52 images, 28 from 5 desmin knock-out (DKO) mice, and 24 images from 5 wild-type (WT). All images combined resulted in 208,270 analyzed muscle cells, with each image taking approximately 2 minutes of computing time for U-Net-Synths predictions and approximately 3 minutes for the analysis pipeline within Fiji for a total of 5 minutes per image. The exact time was dependent on the individual image in regards to image size (resolution) and the amount of muscle cells it contained. This resulted in the fully automated approach being approximately 360 times faster than the manual labeling approach. Based on the mean muscle cell diameter, i.e. the mininmum Feret diameter (MFD), and the standard deviation in addition to the muscle cell area, i.e. cross-sectional area (CSA) and its standard deviation, the results showed two visibly distinct clusters consisting of solely DKO and WT samples respectively, which could have been separated with a linear boundary. The results from the connective tissue analysis also showed a statistical difference large enough to form two visually separate clusters for both DKO and WT samples for the whole slides and the soleus muscle. Additionally, the fully automated nature of the presented approach allowed us to analyze the gastrocnemius muscle (GM) and soleus muscle (SM) separately in addition to the whole muscle sample. In addition to the novel connective tissue analysis, we were able to gain two additional novel insights into the DKO mouse model from the separated results. First, the GM showed a change that again created two separate clusters for both DKO and WT samples by just analyzing the muscle cells from the GM. We found that there are enlarged cells in the GM in addition to the more commonly described muscle cells with reduced size in the SM. We assume that the change comes from the muscle cells in the GM to perform more myofibrillogenesis to compensate for the weakened SM. Secondly, visualization of the results in the GM suggests that the DKO phenotype is dependent on the exact location within the posterior compartment of the leg. The results indicate a compartment-based difference in the phenotype, but further experiments are necessary to confirm these findings. In conclusion, the fully automatic muscle cell analysis method presented in this thesis did not just increase the speed of skeletal muscle sample analysis in comparison to the state-of-the-art method, but it also allowed to analyze a larger amount of muscle cells in DKO samples, which revealed novel insights into the mouse model and potentially the anatomy of the calf muscle in mice.
Oliver Aust (Fri,) studied this question.