Abstract
UNet segmentation models have demonstrated superior performance compared to conventional handcrafted features. Modality specific DL models are better at transferring domain knowledge to a relevant target task than those that are pretrained on stock photography images. Using them helps improve model adaptation, generalization, and class-specific region of interest localization. In this study, we train custom chest X ray modality specific UNet models for semantic segmentation of Tuberculosis (TB) consistent findings. Automated segmentation of such manifestations could help radiologists reduce errors following initial interpretation and before finalizing the report. This could improve radiologist accuracy by supplementing decision making while improving patient care and productivity. Our approach uses a comprehensive strategy that first uses publicly available chest X ray datasets with weak TB annotations, typically provided as bounding boxes, to train a set of UNet models. Next, we improve the results of the best performing model using an augmented training strategy on data with weak localizations from the outputs of a selection of DL classifiers that are trained to produce a binary decision ROI mask for suspected TB manifestations. The augmentation aims to improve performance with test data derived from the same training distribution and other cross institutional collections. We observe that compared to non augmented training our augmented training strategy helped the custom modality specific UNet models achieve superior performance with test data that is both similar to the training distribution as well as for cross institutional test sets.
Abstract (translated)
URL
https://arxiv.org/abs/2102.10607