Improved Semantic Segmentation of Tuberculosis-Consistent Findings in Chest X-rays Using Augmented Training of Modality-Specific U-Net Models with Weak Localizations.

Sivaramakrishnan Rajaraman,Sameer K Antani,Philip O Alderson,Les R Folio,Jane Dimperio

doi:10.3390/diagnostics11040616

Abstract

Deep learning (DL) has drawn tremendous attention for object localization and recognition in both natural and medical images. U-Net segmentation models have demonstrated superior performance compared to conventional hand-crafted feature-based methods. Medical image modality-specific DL models are better at transferring domain knowledge to a relevant target task than those pretrained on stock photography images. This character helps improve model adaptation, generalization, and class-specific region of interest (ROI) localization. In this study, we train chest X-ray (CXR) modality-specific U-Nets and other state-of-the-art U-Net models for semantic segmentation of tuberculosis (TB)-consistent findings. Automated segmentation of such manifestations could help radiologists reduce errors and supplement decision-making while improving patient care and productivity. Our approach uses the publicly available TBX11K CXR dataset with weak TB annotations, typically provided as bounding boxes, to train a set of U-Net models. Next, we improve the results by augmenting the training data with weak localization, postprocessed into an ROI mask, from a DL classifier trained to classify CXRs as showing normal lungs or suspected TB manifestations. Test data are individually derived from the TBX11K CXR training distribution and other cross-institutional collections, including the Shenzhen TB and Montgomery TB CXR datasets. We observe that our augmented training strategy helped the CXR modality-specific U-Net models achieve superior performance with test data derived from the TBX11K CXR training distribution and cross-institutional collections (p < 0.05). We believe that this is the first study to i) use CXR modality-specific U-Nets for semantic segmentation of TB-consistent ROIs and ii) evaluate the segmentation performance while augmenting the training data with weak TB-consistent localizations.

Highlights

Tuberculosis (TB) is caused by Mycobacterium tuberculosis and most often affects the lungs but can occur throughout the body
The materials and methods are further divided into the following subsections: (i) dataset characteristics; (ii) statistical analysis; (iii) chest X-ray (CXR) modality-specific pretraining; (iv) lung segmentation and preprocessing; (v) model fine-tuning and weak TB-consistent localization; (vi) verifying ROI localization; (vii) TB-consistent ROI segmentation using U-Net models; (viii) selecting appropriate loss function and other evaluation metrics; (ix) reducing interobserver variability using simultaneous truth and performance level estimation (STAPLE)-based consensus ROI generation, and (x) task-appropriate data augmentation
The results are organized under the following sections: (i) CXR modality-specific pretraining; (ii) lung segmentation; (iii) fine-tuning toward TB detection; (iv) TB-consistent ROI localization and mask generation; (v) TB-consistent ROI segmentation evaluation using TBX11K train/test dataset, and (vi) TB-consistent ROI segmentation evaluation using TBX11K train set and cross-institutional test sets

Summary

Introduction

Tuberculosis (TB) is caused by Mycobacterium tuberculosis and most often affects the lungs but can occur throughout the body. Automated techniques have been proposed to offset some of these challenges It is unclear if these methods are generalizable across different datasets or concur with radiologist findings. These concerns can be addressed using automated segmentation methods that determine a prediction model to obtain highly accurate results over the training dataset while maintaining good generalization for crossinstitutional test sets. We are challenged by issues concerning varying image acquisition, the variability with which the TB-consistent findings may manifest in the images, and insufficiently or weakly labeled publicly available data sets. We further discuss (i) the need for automated segmentation of TB-consistent findings, (ii) the need to train medical modality-specific segmentation models, and (iii) the methods proposed in this study to improve automated segmentation of TB-consistent findings in CXRs

Methods

Results

Conclusion