[Paper Review] Learning to recognize Abnormalities in Chest X-Rays with Location-Aware Dense Networks

Sebastian Guendel, Saša Grbić+5 more|arXiv (Cornell University)|Mar 12, 2018

COVID-19 diagnosis using AI7 references40 citations

TL;DR

This paper introduces DNetLoc, a location-aware DenseNet that uses high-resolution chest X-ray images and spatial pathology information, achieving state-of-the-art AUC on ChestX-Ray14 and strong PLCO results with patient-wise splits.

ABSTRACT

Chest X-ray is the most common medical imaging exam used to assess multiple pathologies. Automated algorithms and tools have the potential to support the reading workflow, improve efficiency, and reduce reading errors. With the availability of large scale data sets, several methods have been proposed to classify pathologies on chest X-ray images. However, most methods report performance based on random image based splitting, ignoring the high probability of the same patient appearing in both training and test set. In addition, most methods fail to explicitly incorporate the spatial information of abnormalities or utilize the high resolution images. We propose a novel approach based on location aware Dense Networks (DNetLoc), whereby we incorporate both high-resolution image data and spatial information for abnormality classification. We evaluate our method on the largest data set reported in the community, containing a total of 86,876 patients and 297,541 chest X-ray images. We achieve (i) the best average AUC score for published training and test splits on the single benchmarking data set (ChestX-Ray14), and (ii) improved AUC scores when the pathology location information is explicitly used. To foster future research we demonstrate the limitations of the current benchmarking setup and provide new reference patient-wise splits for the used data sets. This could support consistent and meaningful benchmarking of future methods on the largest publicly available data sets.

Motivation & Objective

• Motivate automated chest X-ray abnormality detection to support radiologists and reduce reading errors.
• Address data leakage concerns by using patient-wise splits and benchmarks.
• Leverage high-resolution inputs and explicit spatial localization to improve classification accuracy.
• Provide new patient-wise benchmarks on large public datasets to enable principled future comparisons.

Proposed method

• Use a variant of DenseNet-121 with sigmoid outputs for multi-label classification.
• Incorporate high-resolution (1024x1024) inputs via two initial strided conv layers initialized to Gaussian down-sampling.
• Apply class-weighted binary cross-entropy to handle label imbalance.
• Integrate spatial location information from PLCO for selected pathologies via additional binary location-labels and cross-entropy losses.
• Pool ChestX-Ray14 and PLCO into a single training framework with 35 labels (14 ChestX-Ray14 + 21 PLCO).
• Train with Adam optimizer (lr=1e-3, reduce on plateau) and a batch size of 128.

Experimental results

Research questions

• RQ1 Does incorporating spatial location information improve pathology classification on chest X-ray images?
• RQ2 What are the benefits of training with high-resolution input for chest X-ray abnormalities?
• RQ3 How does pooling data from ChestX-Ray14 and PLCO affect performance and benchmarking?
• RQ4 Can a patient-wise split setup provide more robust and fair benchmarking than random-image splits?
• RQ5 To what extent can location-aware modeling reduce training time while improving accuracy?

Key findings

• On ChestX-Ray14, the proposed approach achieves the best average AUC on the official test split (mean AUC around 0.807 for the DNet baseline; the novel split yielded mean AUC of 0.841 with 0.019 SD).
• Using location information on the PLCO dataset yields higher AUC for pathologies with spatial labels (average gain ~2.3%).
• Final PLCO mean AUC reaches 0.874 when training on the combined ChestX-Ray14 and PLCO data with location information.
• Spatial labeling improves per-pathology AUCs such as Nodule, Mass, Granuloma, Infiltrate, Scaring, Fibrosis, Bone/Soft Tissue Lesion, Cardiac Abnormality, and Effusion.
• Training time is reduced by about a factor of 2 when location information is used.
• The study highlights limitations of previous benchmarks and provides new patient-wise splits for principled comparison.