Functional tissue monitoring: Intraoperative estimation of tissue parameters such as oxygenation, enabling the objective identification of perfused versus ischemic tissue. Tissue differentiation: Since distinguishing tissue types can be difficult — even for experienced surgeons — her work explored automated semantic segmentation of surgical scenes. Sepsis diagnosis: As diagnosing postoperative sepsis is often slow and inconclusive and mortality risk increases with every hour of delayed treatment, Silvia investigated rapid and reliable detection of sepsis in postoperative intensive care unit (ICU) patients. To address these challenges, Silvia used hyperspectral imaging (HSI), a technology originally developed for remote sensing. Changes in tissue composition, such as differences in perfusion, tissue type, or early signs of sepsis (e.g., edema and microcirculatory dysfunction), alter the spectral signatures of biological tissue. However, these changes often remain invisible to the human eye. This limitation arises because human vision (and conventional RGB cameras that mimic it) captures only three broad color channels: red, green, and blue. HSI overcomes this restriction by measuring light in many narrow wavelength bands, often extending beyond the visible spectrum, and thus capturing subtle reflectance differences at every pixel. Because HSI data is high dimensional, deep learning is well suited to identify patterns related to tissue physiology and disease. Silvia demonstrated that HSI-based deep learning models for surgical scene segmentation can achieve performance comparable to a human expert. In what is currently the largest clinical HSI study, she further showed that deep learning models can accurately distinguish septic from non-septic patients in a surgical ICU and outperform widely used biomarkers and clinical scores, while enabling rapid, non-invasive, cost-effective and mobile assessments. A major barrier to the clinical deployment of AI systems is domain shift, which can severely degrade model performance. While increasingly studied in computer vision, this problem has received little attention in medical HSI. Silvia addressed this gap by systematically analyzing real-world domain shifts related to hardware, scene geometries and populations. To support further research and clinical translation in this emerging field, Silvia also released public datasets as well as code and pretrained models. Her full thesis is available here. 23 DAILY WACV Monday Silvia Seidlitz
RkJQdWJsaXNoZXIy NTc3NzU=