Computer Vision News 26 Congrats, Doctor Mumu! First, she developed a quantitative collateral scoring method employing low-rank and sparse decomposition, focusing on radiologically assessing filled versus unfilled vessels over time using 4D CTA. Further, recognizing the potential of deep learning over the traditional methods, in automating feature extraction, she implemented a deep learningdriven automatic evaluation system using 4D CTA. This approach leverages knowledge transfer from a pre-trained EfficientNet B0 network (Figure below for the architecture used in this approach) to alleviate the substantial manual engineering typically associated with classical machine learning and quantitative methods. This method utilized focal loss and 2D MIPs from 4D CTA to alleviate the imbalanced and small dataset issues. Given that radiologists compare an ischemic patient's affected and unaffected sides to determine collateral scores, Mumu further introduced an approach following this criteria that enhances the efficacy of automated evaluation through machine learning. This method focuses on the radiomic features extracted from ischemic damage using NCCT images of both sides of the brain to evaluate collaterals (figure below). Further, to validate the previous method, a technique employing Siamese networks is proposed which addresses the challenges of small and imbalanced NCCT datasets in collateral evaluation. This approach enhances adaptability to data-scarce medical tasks. To enhance collateral evaluation, Mumu finally introduced a few-shot learning-based 3D blood vessel segmentation approach, minimizing the need for extensive label annotation in time-consuming 3D scenarios and overcoming limited pre-trained weights for deep learning
RkJQdWJsaXNoZXIy NTc3NzU=