A new chest X-ray retrieval model reports better results than global and local comparison systems, with gains of 18% to 26% in retrieval accuracy and 11% to 23% in ranking quality. The work is an arXiv preprint dated 28 August 2026.
The system, called CheXtriev, targets visual-only case-based retrieval. Given a chest radiograph, it searches for relevant radiographs, and the study asks whether an anatomy-aware representation can make that search better.
A search built around anatomy
The approach starts with anatomically defined regions rather than treating the image as a single undifferentiated object. It represents those regions in a graph, where the regions act as nodes and the connections capture spatial relationships and co-occurrence between them.
CheXtriev adds learnable location and edge embeddings, global edge-aware attention and gated residuals. Together, these components are designed to preserve information about both individual regions and the way they relate across the radiograph.
Each image is converted into a dense anatomy-aware embedding, or numerical representation, and searched through a fast, exact FAISS k-nearest-neighbor index. The system uses inner-product and cosine similarity to return the closest visual matches.
What the benchmark found
The evaluation used frontal PA or AP radiographs from MIMIC-CXR-JPG v2.0.0 and the dataset’s official splits. The selected images had valid bounding boxes for all eighteen anatomical regions. The split contained 226,473 training images, 1,863 validation images and 3,191 testing images.
Nine findings were evaluated. A radiograph counted as positive when any region carried a positive label, and each test image was queried against the remaining test set.
The authors summarized retrieval with Average Precision, or AP, for ranking quality, Hit Ratio, or HR, for retrieval effectiveness, and Reciprocal Rank, or RR, for how early a relevant result appears. The main comparison examined the top five retrieved images.
For top-five retrieval, CheXtriev had mean AP of 55.1, mean HR of 40.7 and mean RR of 57.4. Its weighted mean values were 70.8 for AP, 59.5 for HR and 73.4 for RR, exceeding the corresponding displayed baseline means.
Against both global baselines, the paper reports higher AP for all nine findings. The reported AP values ranged from 91.7% for LO to 28.8% for FO/HF, while reported mean-AP improvements over global approaches ranged from at least 12% to 23%.
For lower-prevalence findings, the reported improvement ranges were 82.3% to 94.6% for FO/HF, 40.6% to 253.9% for PTX, 30.6% to 35.6% for CONS and 14.6% to 20.5% for PN.
What the design tests suggest
An internal comparison also favored the fuller regional design over a simple mean-pooled baseline. V0 scored 51.8 mAP, 39.7 mHR and 54.0 mRR, while V6 scored 55.1 mAP, 40.7 mHR and 57.4 mRR.
The comparison with the local AnaXNet baseline showed higher mean AP for several examples. The reported increases were 61.80% for FO/HF, 16.82% for PE/HO and 14.49% for CONS.
Other variant comparisons were less uniform. V4 was reported as 6.38% lower in mAP and 6.89% lower in mRR than V6, while V6 and V3 scored above V5 and V2, respectively. The supplied analysis does not give exact inferential statistics for these contrasts.
A benchmark result with clear boundaries
The headline percentage ranges need qualification. The abstract does not map each range to particular findings or comparators, and confidence intervals are not reported. Some table entries are marked p < 0.05, but the exact p-values are not supplied.
The data scope is also specific: the evaluation covers frontal PA or AP images from MIMIC-CXR-JPG v2.0.0 with valid boxes for all eighteen regions. The scores therefore describe performance on this selected retrieval benchmark.
The interpretability evidence is preliminary. In a single ECS query, the first three CheXtriev retrievals matched ECS and their saliency maps focused on the cardiac silhouette. The fourth retrieval was incorrect, with attention diffused through the lung region. That example illustrates the approach but is not a broad validation of the saliency maps.
The supplied record identifies the manuscript as arXiv:2608.28137v1, dated 28 August 2026. It reports no funding source, acknowledges Dr. L.T. Kishore for radiograph-interpretation insights and states that the authors have no conflicts of interest.
Paper data and sources
Original title: CheXtriev: Anatomy-Centered Representation for Case-Based Retrieval of Chest Radiographs
Authors: Naren Akash, Arihanth Tadanki, Jayanthi Sivaswamy
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text