Source-linked AI summary
Learning to recognize Abnormalities in Chest X-Rays with Location-Aware Dense Networks
Sebastian Guendel, Sasa Grbic, Bogdan Georgescu, Kevin Zhou, Ludwig Ritschl, Andreas Meier, Dorin Comaniciu
TL;DR
Chest X-ray abnormality classifiers often use random image splits and omit spatial or high-resolution information, complicating reliable evaluation. The paper proposes DNetLoc, which jointly uses high-resolution images and pathology-location labels, and reports improved performance when location information is available. It also provides patient-wise benchmark splits for future evaluation.
Problem
Random image-based splitting can place the same patient in training and test sets, while existing methods often omit pathology location and high-resolution image information.
Method
DNetLoc combines high-resolution chest X-ray processing with spatial pathology labels and trains across the ChestX-Ray14 and PLCO datasets.
Results
For PLCO, location information improved performance by an average of 2.3% for pathologies with such labels, while the final mean AUC score was 87.4%.
Takeaways & Limitations
The paper provides patient-wise splits and a benchmarking setup intended to support principled comparison of future chest X-ray abnormality-detection methods.
Takeaways & Limitations
ChestX-Ray14 follow-up studies can have incomplete and inconsistent labeling because reports often focus on specific clinical questions.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper addresses chest X-ray abnormality detection under flawed random image splits and limited use of spatial and high-resolution information. It proposes DNetLoc and patient-wise benchmarking to improve evaluation and classification.
- Research gap: Random image splits can place the same patient in training and test sets because ChestX-Ray14 averages 3.6 images per patient.This also creates variability in classification performance between splits.
- Proposed approach: DNetLoc incorporates pathology location information and high-resolution X-ray data for abnormality classification.The method is evaluated using ChestX-Ray14 and PLCO.
- Evaluation: The study evaluates 86,876 patients and about 297,541 images across the two datasets.The authors describe this as the largest dataset reported in the community.
- Results and benchmarking: The authors report the best performance on the existing ChestX-Ray14 benchmark with published patient-wise training and test splits.They also propose a new benchmarking setup with patient-wise splits for future comparisons.
2 Datasets
The study combines two chest X-ray datasets with different sizes, image formats, labels, and spatial annotations. Their pathology labels are imbalanced, and the authors keep similarly named labels separate because definitions may differ.
- Dataset composition: ChestX-Ray14 contains 30,805 patients, 112,120 images, 1024 × 1024 8-bit grayscale images, and 14 pathology classes.
- Dataset composition: PLCO contains 56,071 patients and 185,421 images originally sized 2500×2100 with 16-bit grayscale values.The experiments use 12 prevalent pathology labels, five of which include spatial information.
- Label harmonization: Six labels share names across the datasets, but the authors treat them as separate because equivalent label definitions cannot be guaranteed.They also assume there is no patient overlap between the datasets.
- Label distribution: Pathology labels are highly imbalanced, posing a challenge to learning algorithms.Figure 1 displays image counts across pathologies while excluding images without any listed pathology.
3 Method
The method uses a weighted multi-label DenseNet architecture that accepts high-resolution images and adds spatial-location supervision when available. Dataset-specific labels and losses allow joint training despite differing annotations.
- 3.1 Multi-label Setup: The ChestX-Ray14 task treats 14 pathology labels independently using sigmoid outputs and weighted binary cross-entropy losses.Weights account for the numbers of presence and absence samples.
- 3.2 Leveraging High-Resolution Images and Spatial Knowledge: Two strided convolutional layers precede DenseNet-121 to process 1024 × 1024 chest X-ray inputs and exploit high-resolution information.The added filters are initialized to Gaussian down-sampling operations.
- 3.2 Leveraging High-Resolution Images and Spatial Knowledge: PLCO spatial supervision encodes lung side, fifth-based within-lung location, and diffuse disease using additional classes.Exact positions of multiple and diffuse diseases are not provided.
- 3.2 Leveraging High-Resolution Images and Spatial Knowledge: Spatial-location labels are trained as independent binary classes with cross-entropy losses, with the number of present labels depending on available disease-location information.
- 3.3 Dataset Pooling: The combined network has 35 labels: 14 from ChestX-Ray14 and 21 from PLCO, with dataset-specific label availability handled during joint training.The architecture routes classes according to the dataset of each sample.
4 Experimental Results
Experiments use patient-wise splits to avoid patient overlap and evaluate DNet/DNetLoc on ChestX-Ray14 and PLCO. Results show improved AUC with the proposed approach, especially when pathology location information is available.
- Evaluation setup: Patient-wise splits allocate 70% of patients for training, 10% for validation, and 20% for testing.The paper avoids random image splits because patients average 3.6 images in ChestX-Ray14 and 3.3 in PLCO, making cross-subset patient overlap likely.
- ChestX-Ray14: 0.841 mean AUC is achieved on the more representative random patient split for ChestX-Ray14.Table 1 contrasts this right-column result with the official ChestX-Ray14 test-set evaluation.
- ChestX-Ray14: The official ChestX-Ray14 evaluation shows the proposed network improves mean AUC by over 5% over previous work.The authors also identify label inconsistency and differing image counts per patient between official training and test sets as limitations of that split.
- PLCO: On PLCO, DNetLoc uses spatial information for five pathologies and achieves the best AUC scores in the table.Both DNet and DNetLoc are trained on the combined ChestX-Ray14 and PLCO datasets.
- ChestX-Ray14: 0.831 average AUC with 0.019 standard deviation is reported across the computed random patient splits.Detailed performance for the novel patient-wise benchmark is shown in Table 1 and Fig. 4 left.
- Limitations: Follow-up exams in ChestX-Ray14 show substantial label variance, producing incomplete labels because labels are generated from reports.Follow-up reports may focus on a specific question, such as whether pneumothorax disappeared, rather than consistently labeling other abnormalities.
- PLCO: 87.4% final mean AUC is reached on PLCO, while location information improves the supported pathologies by an average of 2.3%.Using location information also reduces training time by a factor of 2.
5 Conclusion
The paper presents location-aware Dense Networks that exploit high-resolution chest X-rays and pathology spatial information. It evaluates them on a large multi-dataset cohort, reports improved classification, and proposes patient-wise benchmarking to address prior validation limitations.
- Conclusion: DNetLoc classifies chest X-ray pathologies by incorporating spatial pathology information and exploiting high-resolution image data.The method is presented as a location-aware Dense Network.
- Conclusion: The system is trained and validated on 86,876 patients and 297,541 images.This is described as the largest chest X-ray data set used in the work.
- Conclusion: Patient-wise splits and a new benchmarking setup are proposed to support principled evaluation of future chest X-ray abnormality-detection methods.The paper also reports limitations in the validation strategy of previous work.
- Scope: The feature is research-only and is not commercially available, with future availability not guaranteed for regulatory reasons.This disclaimer bounds the current deployment scope.