Lung lobe segmentation in chest CT is relevant to a wide range of clinical applications. However, existing segmentation pipelines often exhibit vulnerabilities and performance degradations when applied to external datasets. This is usually attributed to the size of the available dataset or model. We show that it is possible to enhance generalizability without huge resources by carefully curating the dataset and combining machine learning with medical expertise. Multiple machine learning techniques (self-supervision (SSL), attention (A), and data augmentation (DA)) are used to train a fast and fully-automated lung lobe segmentation model based on 2D U-Net. Our study involved evaluating these techniques on a diverse dataset collected under the RACOON project, encompassing 100 CT chest scans from patients with bacterial, viral, or SARS-CoV2 infections. We compare our model to a baseline U-Net trained on the same dataset. Our approach significantly improved segmentation accuracy (Dice score of 92.8% vs. 82.3%, p < 0.001). Moreover, our model achieved state-of-the-art performance (Dice score of 92.8% vs. 90.8% for the literature’s state-of-the-art, p = 0.102) with reduced training examples (69 vs. 231 CT Scans). Among the techniques, data augmentation with expert knowledge displayed the most significant impact, enhancing the Dice score by +0.056. Notably, these enhancements are not limited to lobe segmentation but can be seamlessly integrated into various medical imaging segmentation tasks, demonstrating their versatility and potential for broader applications.
Identification of cracks is essential to assess the structural integrity of concrete infrastructure. However, robust crack segmentation remains a challenging task for computer vision systems due to the diverse appearance of concrete surfaces, variable lighting and weather conditions, and the overlapping of different defects. In particular recent data-driven methods struggle with the limited availability of data, the fine-grained and time-consuming nature of crack annotation, and face subsequent difficulty in generalizing to out-of-distribution samples. In this work, we move past these challenges in a two-fold way. We introduce a high-fidelity crack graphics simulator based on fractals and a corresponding fully-annotated crack dataset. We then complement the latter with a system that learns generalizable representations from simulation, by leveraging both a pointwise mutual information estimate along with adaptive instance normalization as inductive biases. Finally, we empirically highlight how different design choices are symbiotic in bridging the simulation to real gap, and ultimately demonstrate that our introduced system can effectively handle real-world crack segmentation.