Organizations: the School of Computer Science, Northwest University, Shaanxi, China · the School of Electronic Information, Northwest University, Shaanxi, China · the School of Art, Northwest University, Shaanxi, China
Radiology report generation (RRG) streamlines clinical workflows and alleviates the burden of radiologists' documentation. Standard vision-language models capture global visual cues and lack fine-grained alignment between anatomical regions and diagnostic text. Most retrieval-augmented methods retrieve complete historical reports, introducing redundant noise and severe vision-text mismatches that trigger clinically risky hallucinations and missed lesions. In this paper, we introduce AR-RRG, an anatomy-guided retrieval-driven method for the generation of radiology reports. AR-RRG consists of two core modules: the anatomy-aware self-training lesion detection network (ASLDNet) and the retrieval-driven radiology report generation (RRRG). ASLDNet leverages anatomy-aware spatial verification (ASV) and class-adaptive dynamic thresholding (CDT) via self-training, delivering accurate lesion localization to support subsequent generation. As the generative branch of AR-RRG, RRRG eliminates clinical hallucinations through dual-granularity retrieval (DR) and complements missing lesions with a neighborhood consensus (NC) mechanism. A parallel continuous visual perception (PCVP) stream converts radiographs into visual tokens through fixed medical encoding and resampling. Multi-scale text and visual features are fused, followed by LoRA-efficient large language model tuning to generate clinically consistent, visually grounded radiology reports. Extensive evaluations on MS-CXR, MIMIC-CXR, and IU X-Ray demonstrate that AR-RRG consistently outperforms state-of-the-art methods across mainstream metrics. Our code will be made publicly available.
Figures & tables
Fig. 1: (a) Per-Class Confidence Distributions. (b) Scatter plot of confidence versus IoU. The yellow area denotes valid lesions filtered out by low confidence, and the red area corresponds to high-confidence false positives. (c) Detection errors. Colored areas mark valid anatomical regions; red dashed boxes show predictions.
Fig. 2: The overall framework of our proposed AR-RRG.
TABLE V: Comparison with state-of-the-art methods on MIMIC-CXR dataset. “-” denotes that the original paper did not report relevant results.
Method
Micro-F1 14
Micro-F1 5
Macro-F1 14
Macro-F1 5
F1-RG
B-1
B-4
RG-L
CEMRAG [ 5 ]
0.498
0.526
0.314
0.443
0.187
0.297
0.061
0.201
ASLDNet (Detection Only)
0.543
0.706
0.163
0.382
0.139
0.269
0.036
0.169
+ Template Retrieval
0.548
0.708
0.172
0.397
0.145
0.271
0.040
0.175
+ Template & Phrase Retrieval
0.526
0.667
0.185
0.386
0.221
0.359
0.078
0.216
+ Template & Phrase + NC
0.582
0.699
0.250
0.478
0.226
0.344
0.074
0.219
AR-RRG (Full)
0.630
0.744
0.348
0.582
0.352
0.425
0.152
0.322
TABLE VI: Ablation studies on MIMIC-CXR. Each subsequent row accumulates all previous components: Detection Only → +Template Retrieval → +Phrase Retrieval → +Neighborhood Consensus (NC). The full model further adds PCVP and LoRA fine-tuning. CEMRAG [ 5 ] is shown as a reference. Best results are in bold .
Label
Δ TP
Δ FP
Precision
Pleural Effusion
+422
+186
69.4%
Support Devices
+294
+167
63.8%
Lung Opacity
+154
+80
65.8%
Edema
+129
+102
55.8%
Cardiomegaly
+62
+66
48.4%
Consolidation
+11
+33
25.0%
TABLE VII: Per-label changes in true positives (TPs) and false positives (FPs) introduced by the Neighborhood Consensus (NC) module on MIMIC-CXR. Δ TP and Δ FP denote the changes after adding NC. Recovery precision is defined as Δ TP/( Δ TP+ Δ FP). “-” indicates that ΔTP+ΔFP≤0 , for which recovery precision is undefined. The Total row reports the sums of Δ TP and Δ FP across all labels.
Method
B-1
B-2
B-3
B-4
MTR
RG-L
R2Gen
0.470
0.304
0.219
0.165
0.187
0.371
CMCL
0.473
0.305
0.217
0.162
0.186
0.378
KM-GSK
0.496
0.327
0.238
0.178
-
0.381
CPLIAT
0.479
0.301
0.213
0.155
0.195
0.363
DCL
-
-
-
0.163
0.193
0.383
R2GenGPT
0.488
0.316
0.228
0.173
0.211
0.377
TABLE VIII: Comparison with state-of-the-art methods on the IU X-Ray dataset. Methods marked with ∗ are evaluated in a zero-shot cross-domain setting (trained on MIMIC-CXR only); all remaining methods are fine-tuned on IU X-Ray.
Fig. 3: Qualitative comparison of lesion detection results on MS-CXR across eight pathology categories. Green boxes denote ground truth; colored boxes denote predictions from different methods.
Fig. 4: Qualitative comparison of report generation on a representative MIMIC-CXR case. The progression from Vision-Only to AR-RRG (Full) demonstrates the effectiveness of each module: detection guidance reduces hallucinations, dual-granularity retrieval improves localization, and neighborhood consensus recovers missing device information. The final model achieves high clinical fidelity with visual feature fine-tuning. Color coding indicates incorrect (red), partially accurate (purple), and accurate (green) content.