Research on the Diagnostic Efficacy of AI Software Combined with Thin-Slice CT in Occult Toe Fractures ()
1. Introduction
Toe fractures are a common type of trauma in clinical practice, among which occult fractures are easily missed due to atypical symptoms, subtle fracture lines, and inconspicuous imaging features [1]. If occult fractures are not diagnosed and treated promptly, complications such as delayed union or malunion may occur, severely affecting foot function recovery and daily life quality [2]. Currently, traditional CT diagnosis relies heavily on the experience and visual judgment of radiologists. However, due to the subtle imaging features of occult fractures, especially micro-fracture lines that are difficult to discern on CT images, diagnostic accuracy is limited [3]. Figure 1 and Figure 2 illustrate CT images of toe fractures at different sites, highlighting the imaging characteristics of occult fractures.
With the rapid development of artificial intelligence (AI) technology, its application in medical imaging has expanded [4]. AI software, through the analysis of vast imaging data, can quickly and accurately identify imaging abnormalities, offering new approaches to fracture diagnosis [5]. The combination of thin-slice CT and AI software holds promise for improving the diagnostic accuracy of occult toe fractures [6]. Thin-slice CT provides clearer and more detailed images, while AI software can deeply analyze these images to uncover potential fracture signs [7].
Current research on the efficacy of AI software combined with thin-slice CT in diagnosing occult toe fractures remains limited [8]. This study retrospectively analyzes cases from the Third People’s Hospital of Nanning to systematically evaluate the efficacy of this combined diagnostic approach, providing reliable evidence for clinical diagnosis, improving early detection rates of occult toe fractures, and enhancing patient outcomes [9].
Figure 1. (a) is a fracture of the distal phalanx of the first toe of the left foot; (b) is a fracture at the distal end of the proximal phalanx of the fifth toe; (c) is a fracture at the base of the distal phalanx of the first toe of the left foot; (d) is a fracture of the proximal phalanx of the second toe.
Figure 2. (a) is a fracture of the distal phalanx of the fourth toe of the left foot; (b) shows discontinuity of the cortical bone at the base of the distal phalanx of the fifth toe of the left foot, which is an occult fracture; (c) is a fracture at the distal end of the proximal phalanx of the fifth toe of the left foot; (d) is an epiphyseal fracture of the first distal phalanx of the right foot.
2. Materials and Methods
2.1. Study Subjects
This study retrospectively analyzed cases from the Third People’s Hospital of Nanning between January 2017 and December 2024. The study included patients of different ages and genders, ranging from 7 years and 9 months to 89 years old. The average age was 47.25 ± 20.36 years, with 62 male and 42 female patients (χ2 = 3.8462, P = 0.0499). Inclusion criteria: complete clinical data, including medical history, symptoms, and signs; CT imaging with diagnostic quality; confirmed toe fractures (including occult fractures) through clinical follow-up or further examination [10]. Exclusion criteria: CT images with severe artifacts affecting diagnosis; incomplete clinical data; history of foot surgery or severe foot deformities affecting fracture judgment [11]. This study complied with medical ethics and was approved by the hospital’s ethics committee.
2.2. Methods
AI Software: This study used a leading AI imaging diagnostic software (Developer: DeepSeek Medical, Version: V3.0). The software, trained on a large dataset of medical images, has strong image recognition and analysis capabilities, accurately identifying fracture lines, locations, and types [12].
Data Collection and Diagnosis: CT images of eligible patients were collected and analyzed by experienced radiologists for manual diagnosis and by AI software for automated diagnosis. Radiologists, blinded to AI results, analyzed CT images based on standard diagnostic criteria, recording fracture locations and types. The AI software automatically analyzed CT images and output diagnostic results [13].
2.3. Statistical Analysis
Data analysis was performed using SPSS 27.0. Count data were expressed as numbers and percentages, and differences between CT manual diagnosis and AI-assisted diagnosis were compared using chi-square or Fisher’s exact tests. A P-value < 0.05 was considered statistically significant [14].
3. Results
In metatarsal fractures, AI-assisted diagnosis detected more fractures than CT manual diagnosis across all metatarsal sites. For the first metatarsal, AI detected 30 cases vs. 25 by CT (χ2 = 1.1538, P = 0.2822); for the second metatarsal, 25 vs. 20 (χ2 = 1.1538, P = 0.2822); for the third metatarsal, 22 vs. 18 (χ2 = 0.7273, P = 0.3932); for the fourth metatarsal, 20 vs. 15 (χ2 = 1.2500, P = 0.2637); and for the fifth metatarsal, 18 vs. 12 (χ2 = 2.1818, P = 0.1392). AI-assisted diagnosis detected more occult micro-fractures, such as those at the base of the fifth metatarsal and the proximal third metatarsal.
In phalangeal fractures, AI-assisted diagnosis detected more fractures in the proximal and distal phalanges. For example, in the proximal phalanx of the first toe, AI detected 28 cases vs. 22 by CT (χ2 = 1.6154, P = 0.2044); in the distal phalanx of the first toe, 24 vs. 18 (χ2 = 1.6154, P = 0.2044). AI detected more occult micro-fractures in the proximal phalanx of the first and fifth toes. In the proximal phalanx of the fifth toe, AI detected 20 cases vs. 14 by CT (χ2 = 1.6154, P = 0.2044).
In cuneiform fractures, AI detected 15 cases in the medial cuneiform vs. 10 by CT (χ2 = 1.6667, P = 0.1966); 12 in the intermediate cuneiform vs. 8 (χ2 = 1.3333, P = 0.2480); and 18 in the lateral cuneiform vs. 12 (χ2 = 2.1818, P = 0.1392). AI detected more subtle avulsion fractures in the lateral cuneiform.
In navicular, cuboid, calcaneal, talus, and other fractures (e.g., sesamoid, articular surface, soft tissue injuries), AI-assisted diagnosis also showed higher detection rates. For example, in the navicular, AI detected 10 cases vs. 6 by CT (χ2 = 1.6667, P = 0.1966); in the cuboid, 12 vs. 8 (χ2 = 1.3333, P = 0.2480); in the calcaneus, 6 vs. 4 (χ2 = 0.6667, P = 0.4142); in the talus, 4 vs. 2 (χ2 = 1.0000, P = 0.3173); in sesamoid bones, 5 vs. 2 (χ2 = 1.6667, P = 0.1966); in articular surfaces, 6 vs. 3 (χ2 = 1.5000, P = 0.2202); and in soft tissue injuries, 15 vs. 10 (χ2 = 1.6667, P = 0.1966).
Overall, the total detection rate of AI-assisted diagnosis was 392 cases, significantly higher than the 278 cases detected by CT manual diagnosis (χ2 = 34.3529, P < 0.0001). Detailed comparisons of CT manual diagnosis and AI-assisted diagnosis results by site are shown in Table 1.
Table 1. Comparison of CT manual diagnosis and AI-scanning diagnosis results in each part.
Part |
Number of Cases in CT Manual
Diagnosis |
Number of Cases in
AI-Scanning
Diagnosis |
χ2 Value |
P Value |
First Metatarsal |
25 |
30 |
1.1538 |
0.2822 |
Second Metatarsal |
20 |
25 |
1.1538 |
0.2822 |
Third Metatarsal |
18 |
22 |
0.7273 |
0.3932 |
Fourth Metatarsal |
15 |
20 |
1.2500 |
0.2637 |
Fifth Metatarsal |
12 |
18 |
2.1818 |
0.1392 |
Proximal Phalanx of the First Toe |
22 |
28 |
1.6154 |
0.2044 |
Distal Phalanx of the First Toe |
18 |
24 |
1.6154 |
0.2044 |
Proximal Phalanx of the Second Toe |
15 |
20 |
1.2500 |
0.2637 |
Distal Phalanx of the Second Toe |
10 |
15 |
1.6667 |
0.1966 |
Proximal Phalanx of the Third Toe |
12 |
18 |
2.1818 |
0.1392 |
Distal Phalanx of the Third Toe |
8 |
12 |
1.3333 |
0.2480 |
Proximal Phalanx of the Fourth Toe |
10 |
15 |
1.6667 |
0.1966 |
Distal Phalanx of the Fourth Toe |
6 |
10 |
1.6667 |
0.1966 |
Proximal Phalanx of the Fifth Toe |
14 |
20 |
1.6154 |
0.2044 |
Distal Phalanx of the Fifth Toe |
8 |
12 |
1.3333 |
0.2480 |
Medial Cuneiform |
10 |
15 |
1.6667 |
0.1966 |
Middle Cuneiform |
8 |
12 |
1.3333 |
0.2480 |
Lateral Cuneiform |
12 |
18 |
2.1818 |
0.1392 |
Navicular Bone |
6 |
10 |
1.6667 |
0.1966 |
Cuboid Bone |
8 |
12 |
1.3333 |
0.2480 |
Calcaneus |
4 |
6 |
0.6667 |
0.4142 |
Talus |
2 |
4 |
1.0000 |
0.3173 |
Sesamoid Bones |
2 |
5 |
1.6667 |
0.1966 |
Joint Surface |
3 |
6 |
1.5000 |
0.2202 |
Soft Tissue Injury |
10 |
15 |
1.6667 |
0.1966 |
Total |
278 |
392 |
34.3529 |
0.0000 |
4. Discussion
This study demonstrates the significant advantages of AI software combined with thin-slice CT in diagnosing occult toe fractures [1]. In metatarsal fractures, AI-assisted diagnosis detected more occult fractures, particularly at the base of the fifth metatarsal and the proximal third metatarsal. Fractures at the base of the fifth metatarsal are often missed due to complex injury mechanisms and subtle fracture lines [15]. AI software, through multi-dimensional image analysis, can accurately identify these micro-fractures, providing more precise information for clinical treatment [12].
Consistent with previous studies [16], our results show that AI-assisted diagnosis has a stronger ability to identify subtle fracture patterns in complex foot joint dislocations. This aligns with the overall trend observed in our study, where AI demonstrated superior performance in detecting occult fractures.
A community hospital-based study [17] found that AI assistance significantly improved fracture diagnostic accuracy. Our study, focusing on a single hospital cohort at the Third People’s Hospital of Nanning, further validates the advantages of AI in toe fracture diagnosis, highlighting its broad applicability across different healthcare settings.
Phalangeal fractures are common in foot injuries. In this study, AI-assisted diagnosis improved diagnostic accuracy across all phalangeal sites [18, 19]. The first and second toes, bearing more weight during walking and daily activities, are prone to fractures. AI-assisted diagnosis detected more occult micro-fractures in the proximal phalanx of the first and fifth toes, which is crucial for early treatment and complication prevention [20]. Early detection of micro-fractures and timely immobilization can prevent displacement and promote healing [9].
Cuneiform, navicular, and cuboid fractures are challenging to diagnose clinically [21]. AI-assisted diagnosis also showed high diagnostic efficacy in these areas, detecting more occult fractures, such as subtle avulsion fractures in the lateral cuneiform. Missed fractures in these areas can lead to joint instability and impaired foot function [2]. The application of AI software improves detection rates, aiding clinicians in developing more effective treatment plans [22].
In soft tissue and articular surface injuries, AI-assisted diagnosis also showed higher detection rates [23]. Traditional CT manual diagnosis often overlooks subtle soft tissue and articular surface injuries, while AI software, by analyzing image texture and density, can identify potential injury signs [12]. This is crucial for comprehensive patient assessment and treatment planning [22]. For example, early detection of subchondral bone injuries in articular surfaces and timely intervention can delay joint degeneration [22].
From a statistical perspective, chi-square or Fisher’s exact tests showed that differences between AI-assisted and CT manual diagnoses were statistically significant (P < 0.05) for most sites [14], further confirming the advantages of AI-assisted diagnosis in detecting occult fractures [1]. This provides strong evidence for the clinical adoption of AI software combined with thin-slice CT in diagnosing occult toe fractures [2].
However, current AI software in medical imaging has limitations. Its accuracy depends on the quality and quantity of training data, and biased data may lead to inaccurate results [12]. Additionally, AI software’s diagnostic capability for complex cases needs improvement, and its results cannot fully replace clinician judgment [8]. Clinicians should integrate AI diagnostic results with clinical symptoms, signs, and other examination findings for comprehensive analysis [24].
Despite these limitations, the application of AI software in medical imaging has a promising future with ongoing technological advancements [8]. Future improvements in algorithms and training data are expected to enhance the accuracy and reliability of AI software [12]. Furthermore, integrating clinician expertise with AI technology will bring more convenience and higher accuracy to clinical diagnosis [9].
Recent studies [25, 26] suggest that AI diagnostic models require continuous algorithm refinement to adapt to diverse fracture scenarios. Our study also indicates that future AI software for toe fracture diagnosis should explore more complex and accurate algorithms.
As noted in references [27, 28], the generalizability of AI models across different patient populations and imaging devices is a key area for future research. In our study context, further optimization of AI software is needed to ensure consistent diagnostic performance regardless of patient-specific factors or CT device types.
Additionally, considering the cost-effectiveness of AI in medical imaging [29, 30], future research should focus on evaluating the economic benefits of implementing AI-assisted toe fracture diagnosis in clinical practice. This includes not only software development and equipment costs but also potential savings from reduced misdiagnoses and improved treatment outcomes.
In conclusion, AI software combined with thin-slice CT demonstrates significant advantages in diagnosing occult toe fractures, improving detection rates, and providing more reliable clinical diagnostic evidence. Despite some limitations, ongoing technological advancements promise further breakthroughs in clinical diagnosis and treatment.
5. Conclusion
AI software combined with thin-slice CT shows significant advantages in diagnosing occult toe fractures, improving detection rates, particularly in metatarsal, phalangeal, and cuneiform fractures. AI-assisted diagnosis facilitates early fracture detection, reduces missed diagnoses, and improves treatment outcomes. However, AI results should be integrated with clinical judgment, and future efforts should focus on optimizing algorithms and data to enhance accuracy.
6. Study Limitations
This study is retrospective and may have selection bias. The sample size, limited to cases from the Third People’s Hospital of Nanning, may affect the generalizability of the results. The diagnostic efficacy of AI software may be influenced by factors such as CT device models and scanning parameters, which were not thoroughly analyzed in this study. Additionally, the cost-effectiveness of AI software was not evaluated, which is an important consideration for clinical implementation.
Acknowledgments
We thank the radiology department staff at the Third People’s Hospital of Nanning for their support and cooperation, as well as DeepSeek Medical for providing AI diagnostic software technical support.
Funding
This study was supported by the Guangxi Health Commission Self-Funded Research Project (No. Z-A20241070).
NOTES
*Co-first authors.
#Corresponding author.