Organizations: Dwarkadas J Sanghvi College of Engineering, Mumbai, India · University of Amsterdam, Amsterdam, Netherlands · National University of Singapore, Singapore · Georgia Institute of Technology, Atlanta, USA
Fracture detection and its clinical interpretability see notable improvements when deep vision models are integrated with agentic AI architectures. While deep learning models achieve high diagnostic performance, their black-box nature limits clinical adoption. We propose FRAC-MAS, an agentic AI system for automated, explainable, and safe bone fracture detection. The framework combines a stacked ensemble of four vision models with conformal prediction to produce statistically grounded differential diagnoses, while a multi-agent workflow performs independent verification, retrieves clinical guidelines, and generates patient-friendly reports. A pipeline-depth ablation study confirms that our multi-agent critic triages 86.6% of cases into a high-confidence auto-confirmed cohort while escalating uncertain cases, outperforming a single-agent baseline. Patient preference studies against Llama, MedGemma, and Gemini further demonstrate significantly more comprehensible clinical reports. These results suggest that integrating multi-agent critics with conformal guarantees enables safer radiology triage while preserving clinician oversight. More broadly, FRAC-MAS demonstrates how cooperative agentic architectures can serve as auditable, human-in-the-loop decision support systems for safety-critical healthcare. Our code is available at https://github.com/hardik1712/FRAC-MAS, and the website is available at https://frac-mas.vercel.app.
Figures & tables
Figure 1: End-to-end system architecture: integrating ensemble inference, Grad-CAM localization, multi-agent reasoning, and conformal prediction for verifiable clinical report generation.
Configuration
Acc. (%)
Conf. Acc. (%)
Escalation Rate (%)
Set Size
Coverage (%)
Know/Educ (%)
Single-agent baseline
91.1
—
100.0
—
—
—
Ensemble only
91.1
—
1 0.0
—
—
—
+Conformal ( α=0.10 )
91.1
—
1 0.0
1.08
92.0
—
+Conformal + Critic
91.1
93.8
13.4
1.08
92.0
—
Full pipeline
91.1
93.8
13.4
1.08
92.0
100.0
Table 1: Pipeline-depth ablation on the 112-sample test set ( α=0.10 , stacking). Conf. Acc. is accuracy on cases not flagged for human escalation. † The Critic uses local MedGemma-4B-IT.
Configuration
Det. Rate
F1
sˉ
MaxViT
0.950
0.974
0.947
HC-CBAM-DenseNet169
0.900
0.947
0.903
RAD-DINO
0.914
0.955
0.901
YOLOv26m-cls
0.921
0.959
0.927
Ensemble (Wt. Avg)
0.929
0.963
0.920
Ensemble (Stacking)
0.964
0.982
0.916
Table 2: Binary fracture detection on the Roboflow dataset (140 images, all positive). Det. Rate = recall; sˉ = mean fracture score.
Class
N
Acc. (%)
Coverage (%)
α=0.05
α=0.10
Comminuted (COM)
17
100.0
100.0
100.0
Greenstick (GRN)
13
100.0
100.0
100.0
Healthy (HLT)
10
100.0
100.0
100.0
Oblique (OBL)
17
70.6
76.5
82.4
Oblique Displaced (OBL-D)
9
100.0
100.0
100.0
Table 3: Per-class ensemble accuracy and conformal coverage on the test set. Coverage failures concentrate among morphologically similar types.
Fracture
Active (%)
Mean Act.
MaxViT
HC
MaxViT
HC
Comminuted
2.4
59.8
0.04
0.55
Oblique Disp.
7.4
53.0
0.23
0.52
Spiral
7.4
66.8
0.14
0.61
Transverse
20.6
58.5
0.27
0.57
Table 4: Grad-CAM attention statistics. Active fraction is the proportion of pixels exceeding 0.3 activation. MaxViT attends to 2%-21% of an image while the HyperColumn model covers 53%-67%.
Dimension
Mean Score
Score ≥ 4 (%)
Fleiss’ κ
Technical Accuracy
4.14
75.0
0.72
Comprehensibility
3.95
68.2
0.65
Table 5: Clinician ratings of the system’s educational outputs (1–5 scale) and inter-rater agreement across three orthopedic raters.