Enhancing medical students’ diagnostic accuracy of infectious keratitis with AI-generated images

Abstract

Background

Developing students’ ability to accurately diagnose various types of keratitis is challenging. This study aims to compare the effectiveness of teaching methods—real cases, artificial intelligence (AI)-generated images, and real medical images—on improving medical students’ diagnostic accuracy of bacterial, fungal, and herpetic keratitis.

Methods

97 consecutive fourth-year medical students who had completed basic ophthalmology educational courses were included. The students were divided into three groups: 30 students in the group (G1) using the real cases for teaching, 37 students in the group (G2) using AI-generated images for teaching, and 30 students in the group (G3) using real medical images for teaching. The G1 group had a 1-hour study session using five real cases of each type of infectious keratitis. The G2 group and the G3 group each experienced a 1-hour image reading sessions using 50 AI-generated or real medical images of each type of infectious keratitis. Diagnostic accuracy for three types of infectious keratitis was assessed via a 30-question test using real patient images, compared before and after teaching interventions.

Results

All teaching methods significantly improved mean overall diagnostic accuracy. The mean accuracy improved from 42.03 to 67.47% in the G1 group, from 42.68 to 71.27% in the G2 group, and from 46.50 to 74.23% in the G3 group, respectively. The mean accuracy improvement was highest in the G2 group (28.43%). There were no statistically significant differences in mean accuracy or accuracy improvement among the 3 groups.

Conclusions

AI-generated images significantly enhance the diagnostic accuracy for infectious keratitis in medical students, performing comparably to traditional case-based teaching and real patient images. This method may standardize and improve clinical ophthalmology training, particularly for conditions with limited educational resources.

Background

Infectious keratitis is one of the leading causes of blindness and visual impairment worldwide, and presents a growing public health burden [1,2,3]. Marked by its rapid progression, infectious keratitis can lead to irreversible corneal impairment if not promptly and accurately diagnosed and treated [4]. Clinically, the diagnosis of infectious keratitis mainly relies on the visual inspection of corneal lesions and manifestations under slit-lamp microscopy, which requires a high level of expertise and experience of ophthalmologists [5, 6]. However, the diverse and overlapped visual characteristics of different microbial infections, such as bacterial, fungal, and herpetic keratitis, make accurate diagnosis challenging [7], even for an experienced corneal specialist. The diagnostic accuracy of infectious keratitis also varies significantly among medical institutions in different regions [3, 8]. On the other hand, in medical education, developing students’ ability to accurately differentiate various types of keratitis is a key aspect of teaching reform, especially when considering the integration of artificial intelligence (AI) into traditional medical education.

Recent AI advancements, particularly in the field of image processing and classification, have shown great potential in enhancing diagnostic accuracy across various ocular conditions, especially for retinal and corneal diseases [3, 9,10,11,12]. Previously, we have developed a sequential-level deep learning model for classifying infectious keratitis, achieving an overall accuracy of 80% [8]. This model was developed using a dataset of 115,408 slit-lamp photographs from 10,609 patients with corneal diseases. By using the same dataset, we recently developed an AI generation model based on Stable Diffusion 1.5 for infectious keratitis images. By adjusting the control conditions of text and image form, this model can generate a controlled variety of microbial infectious keratitis images as required. This AI model serves as an educational resource that produces a sufficient number of typical infectious keratitis images, addressing the challenge of limited real-world cases due to the variety of disease presentations and patient morphologies [13].

We hypothesize that training medical students with typical infectious keratitis images can enable them to achieve the same image diagnostic ability as the previous state-of-the-art AI classification model, and that the teaching effect of AI-generated images can match that of real medical images. This study aims to evaluate and compare the effectiveness of teaching methods—real cases, AI-generated images, and real medical images—on improving medical students’ diagnostic accuracy of bacterial, fungal, and herpetic keratitis. This study seeks to provide evidence on the effectiveness of AI-generated images as a teaching tool in ophthalmology education.

Subjects and teaching groups

The study was conducted from April 2024 to December 2024. Ninety-seven consecutive fourth-year medical students of Zhejiang University School of Medicine who had completed the basic ophthalmology educational courses and started the clinical clerkship were enrolled. The lecture on the corneal disease section was taught by a same corneal expert (YFY), and was completed at least one week before the clerkship at the Department of Ophthalmology, Sir Run Run Shaw Hospital. During clerkship, the students were divided into three groups based on the following clinical teaching methods they received: 30 students in the group (G1) using the real cases for teaching, 37 students in the group (G2) using AI-generated images for teaching, and 30 students in the group (G3) using real medical images for teaching.

For the G1 group, a 1-hour clinical case study session was conducted by the preceptor (WX): the preceptor explained five typical real cases of each type of bacterial, fungal, and herpes simplex keratitis by providing the medical history, diagnostic and treatment process, and slit-lamp photographs throughout the course of the disease. Each case was presented with 3–5 images showing the initial phase, post-treatment phase, and final outcomes. For the G2 group, a 1-hour image reading session was conducted under the guidance of the preceptor (WX): the preceptor presented 50 AI-generated images (without medical histories) of each type of bacterial, fungal, and herpes simplex keratitis, with students focusing on independently extracting and analyzing visual features from the images and the preceptor providing answers to questions if any arose. For the G3 group, similar to the G2 group, a 1-hour image reading session was conducted under the guidance of the same preceptor (WX): the preceptor presented 50 slit-lamp photographs each for bacterial, fungal, and herpes simplex keratitis from real patients (without corresponding medical records). In G2 and G3 groups, each presented image represented an independent keratitis case.

AI image generation model

Previous medical image generation tasks typically employed Generative Adversarial Networks (GAN) [14] models as the base of the generative model. GAN models are a type of generative adversarial model that enhances the realism of generated images by adversarial training for a generator and a discriminator to distinguish between real and fake images. Diffusion [15] models represent a novel class of generative models that generate images through denoising and possess a larger number of model parameters, enabling the production of higher-quality images. With the introduction of the Stable Diffusion [16] model, a text conditioning module was incorporated, allowing for controllable generation.

In this study, we adopt a multi-condition diffusion model, generating both text and images simultaneously as conditions for the diffusion model. To this end, we introduced a text-condition module and an image-condition module on the basis of the diffusion model, and combined the two as conditional inputs to the diffusion model (Fig. 1). Additionally, to enhance the representation ability of the generation model for infectious keratitis images (Fig. 2), we fine-tuned the parameters of the two conditional models based on the model parameters of Stable Diffusion 1.5. All AI-generated images used in this study were evaluated by three senior ophthalmologists to verify their clinical accuracy, including both disease classification and representation of disease-specific visual characteristics.

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

Dataset for AI training and image-based teaching

The dataset was selected from 115,408 slit-lamp photographs of 10,609 patients with corneal diseases. Each selected case in the dataset was definitively diagnosed through verification with at least one of the following clinical pieces of evidence: either confirmed by the infection’s resolution through pathogen-specific monotherapy or combination therapy, or established through laboratory identification - with bacterial/fungal infections verified by microscopic smear examination or microbial culture, and viral infections confirmed by polymerase chain reaction testing of tear fluid or corneal scrapings [8]. The final dataset comprised 2153 images (including 437 images of bacterial keratitis, 707 images of fungal keratitis, and 1,009 images of herpes simplex virus stromal keratitis) for AI training and image-based teaching in G3 group.

Image-based diagnostic test and teaching process

A diagnostic test was conducted to evaluate the students’ ability to diagnose infectious keratitis before and after the clinical teaching interventions. The test consisted of 30 single-choice questions, with an equal distribution across three categories: bacterial, fungal, and herpes simplex keratitis, each represented by 10 questions. Students were unaware of the proportion of questions per disease type. Each question presented a slit-lamp photograph from a real patient (which did not appear in any group’s teaching process), and students were required to select one of the three diagnostic options (bacterial, fungal, or viral keratitis) based on their interpretation of the image. All students were tested once before the clinical teaching, without disclosure of the answers. Subsequently, the teaching activity was initiated, and the same set of questions was administered following the completion of the teaching session. Although there was no time limit for the test, all students were able to complete each question within 10 s, both before and after the clinical teaching session. The test scores (represented by accuracy, the proportion of true results among the total number of cases examined [11]) before and after the teaching interventions were used to conduct within-group and between-group comparisons to assess the effectiveness of the three teaching methods.

State-of-the-art AI classification model for comparison

Our previously published state-of-the-art AI classification model, which demonstrated outstanding performance in diagnosing infectious keratitis, was evaluated using the same test administered to the medical students. This model was based on a feature learning mechanism to identify the informative visual patterns of corneal lesions via sequential-level feature learning, that the sampled patches from the center to the edge of the lesion in the corneal image were grouped into a sequential ordered set and fed into a convolutional neural network for feature learning. This model achieved a diagnostic accuracy of 80.00% in a previous testing set [8].

Statistical analysis

Sample size was calculated using the Power Analysis and Sample Size (PASS version 11.0.7; NCSS, LLC. Kaysville, Utah, USA). Data were analyzed using the Statistical Package for Social Sciences (SPSS version 19.0; Cary, NC, USA). The results are expressed as means ± SDs. The Kolmogorov-Smirnov test was applied to test the normality of the data. Paired t-test was used to compare the accuracies before and after the teaching session within each group. One-way analysis of variance (ANOVA) with least significant difference (LSD) and post hoc tests were used for comparing differences in accuracy between different groups. The effect sizes and confidence intervals (CI) were calculated using online Effect Size Calculator (https://www.campbellcollaboration.org/calculator). Correlations between accuracies were tested using the Pearson’s correlation coefficient. The statistical significance level was set at P < 0.05.

Results

The sample size was calculated with an expected 1% difference in diagnostic accuracy between the groups, with a power of 95%, and α = 0.05, the analysis indicated 335 subjects per group. However, due to the limited enrollment within a single academic semester, a total of 97 fourth-year medical students (49 males and 48 females) were included in this study. The performances of the three groups before and after teaching interventions are reported in Table 1; Fig. 3. The state-of-the-art AI classification model achieved a total accuracy of 70.00%. Before teaching, there were no statistical differences in the baseline accuracies between the three groups (ANOVA and post hoc tests, Cohen’s d = 0.1893, 95% CI -0.2421 to 0.6207). After teaching, there were also no statistical differences in accuracy among the three groups (ANOVA and post hoc tests, Cohen’s d = 0.3481, 95% CI -0.0853 to 0.7814).

[IMAGE OMITTED: SEE PDF]

[IMAGE OMITTED: SEE PDF]

All three teaching methods significantly improved the mean overall diagnostic accuracy, as well as the mean accuracies for each type of keratitis (all P < 0.001, paired t-test). Ninety-one (93.81%) students showed improved overall accuracies after the teaching, with the highest improvement being 63% in the G2 group. The mean improvement of overall accuracies was also highest in the G2 group (28.43% ± 16.42%), but there were no statistical differences between the three groups (ANOVA and post hoc tests, Cohen’s d = 0.1578, 95% CI -0.2733 to 0.589). In each of the three groups, there were two cases of overall accuracy decline (accounting for 6.67%, 5.41%, and 6.67%, respectively), with the largest decrease being 10% in the G2 group.

After teaching, the mean accuracy of bacterial keratitis in the G2 group (65.68% ± 18.79%) was significantly lower than the G3 group (79.33% ± 18.93%) (P = 0.002, post hoc test), but the improvement of accuracy was not significantly different (23.51% ± 22.76% and 27.33% ± 27.03%, respectively). In the G2 group, the mean accuracy for bacterial keratitis after teaching (65.68% ± 18.79%) was significantly lower than that for viral keratitis (75.68% ± 16.08%) (P = 0.024, post hoc tests). However, there was no significant difference in the improvement of accuracy for bacterial (23.51% ± 22.76%) and viral (33.24% ± 24.04%) keratitis.

In the G1 group, there was no significant correlation between pre- and post-teaching accuracies. In the G2 group, there was a significant positive correlation between pre- and post-teaching accuracies in fungal keratitis (R = 0.408, P = 0.012). In the G3 group, there was a significant positive correlation between pre- and post-teaching accuracies in viral keratitis (R = 0.539, P = 0.002).

Discussion

This study demonstrated that medical students’ ability to analyze the visual characteristics of infectious keratitis remained insufficient after completing basic ophthalmology educational courses of class lecture. The supplement of real-world training is essential in medical education, and typical case learning is a common teaching method to foster the diagnostic and clinical reasoning skills [17, 18]. However, for infectious keratitis, the condition of real patients is complex and variable, especially with significant variability in visual characteristics of corneal lesions and manifestations caused by different pathogens [2]. Furthermore, encountering a sufficient number of typical cases in clinical practice requires time and a certain degree of coincidence [19]. Moreover, there are differences in teaching experience and capability among different medical institutions and clinical teachers [20]. Hence, creating a more efficient, homogeneous, and stable practical learning environment is warranted to reduce the uncertainty in clinical learning for medical students [21].

The integration of AI-generated images into medical education is a promising strategy to bridge the gap between basic knowledge learning and clinical learning. The model used in this study can accurately and controllably generate images of different types of infectious keratitis, which has the potential to create a more standardized and consistent teaching environment. By generating a wide range of images on demand, AI can provide students with a more comprehensive and varied learning experience, enabling them to better understand and manage the complexity and variability of real-world clinical scenarios. Our study demonstrated that during the clerkship phase, all three teaching methods could enhance diagnostic accuracy of infectious keratitis. Although there was a significant disparity in case numbers among the three groups, this was designed to demonstrate different teaching strategies. G1 group employed the traditional clinical teaching approach, which involved demonstrating and explaining complete treatment processes of typical cases. Due to time constraints, only 5 representative cases for each of the three types of keratitis could be covered. In contrast, G2 and G3 groups were presented solely with keratitis images rather than complete case histories, enabling the display of 50 images per keratitis type within the same teaching timeframe. The results suggested that for diseases where diagnosis relies primarily on visual characteristics, the image-based teaching method (high-load pure visual feature learning) might be more effective than traditional case teaching.

The teaching effectiveness of AI-generated images was comparable to that of real medical images, while sufficient high-quality real medical images with correct annotations are not always available in many medical institutions [4]. The advantages of using AI-generated images for teaching include helping to reduce the variation in clinical teachers’ ability to convey knowledge of infectious keratitis, as well as overcoming the differences in the number, quality, and accessibility of real medical image collections among different medical institutions.

In addition to the inherent difficulty in acquisition, the available corneal datasets frequently exhibit an imbalance among different disease categories and are often contaminated with a large number of non-target diseases [4]. It is particularly difficult to gather comprehensive image sets that fully demonstrate the entire disease progression. Using AI to display the disease process images can serve as a supplement to the existing images. This supplementation is not only used to augment the dataset for training AI [22, 23], but also for the training of medical students. In the future, by more systematically generating images through AI, such as incorporating accurately expressed textual supplements as educational tools, may promote a paradigm shift in medical education. More importantly, the application of AI technology may facilitate the transition of clinical teaching from the traditional teacher-centered model to a new model of human-computer collaboration. This shift may also provide medical students with more opportunities for self-directed learning with AI assisting to achieve self-improvement more efficiently. Therefore, the use of AI-generated images in future medical education offers three advantages: customized disease image generation for targeted medical training, providing a homogeneous and stable teaching environment, and expansion beyond the limitations of existing real image collections.

In this study, medical students who learned through AI-generated images outperformed our previously reported AI classification model on the same test dataset (71.27% vs. 70.00%). Although our previous classification model had an accuracy of 80% on the test set during the developing phase, the accuracy dropped to 70% when changing the test set in this study. This also suggests that humans may have a better ability to generalize the knowledge they have previously learned. Similarly, in a previous teaching study on retinal diseases using AI-synthesized images, it was found that students who learned from feature images synthesized by AI of a certain device could also accurately diagnose images produced by another similar device, demonstrating stronger generalization learning ability compared to AI models [13]. These results suggest that medical AI should focus on enhancing human skills as its development direction, and the future of medical practice will be human-computer collaboration.

There are several limitations to this study. Firstly, the selection of 150 real medical images, 150 AI-generated images, and 30 test images was subjective. The teaching process, including case study guided by teachers and image interpretation by students, were also subjective. The study was inevitably influenced by the teaching proficiency of the preceptor and the students’ ability to extract and summarize visual characteristics. Secondly, AI-generated images still require manual review and affirmation by teachers before use, and the teacher’s guidance is necessary when students are interpreting the images to avoid potential misguidance from AI-generated images. Thirdly, the teaching subjects were students attending clerkship continuously and in batches, making it impossible to randomly assign them into different groups. The three groups of students had the same teacher and content for the basic ophthalmology educational courses, the same time interval between the theory course and the clerkship, and no statistically significant differences in diagnostic accuracy before the teaching interventions, ensuring balanced baseline characteristics among the three groups. The initial power analysis suggested a sample size of 335 participants per group to detect a 1% difference in diagnostic accuracy with 95% power. However, only 97 students were enrolled due to the constraints of a single academic semester. Consequently, the results should be interpreted with caution, and future studies with larger cohorts are needed to validate these findings. Finally, the test was conducted immediately after teaching interventions, which might only reflect the short-term visual feature memory. Since the retention of knowledge achieved in medical learning may diminish over time [24], further research is needed to investigate the long-term effect of the teaching methods, and how well the diagnostic skills learned translate into actual patient care. For example, these students can be recalled after a certain period of time to be re-evaluated using the same diagnostic test, or alternatively, their diagnostic ability can be evaluated when facing real-world patients.

Conclusions

By incorporating AI-generated medical images into the educational curriculum, we have successfully improved students’ diagnostic abilities for diagnosis of infectious keratitis, and it was compatible to the teaching methods with real cases and real medical images. We plan to develop a stable AI-generated corneal disease image model and integrate it with more factors such as medical history, lesion size, response to treatment, and progression of the diseases. Additionally, an interactive human-computer collaborative system is under developing, which can enable medical teachers and students to freely generate corneal images for learning.

Data availability

The datasets generated and/or analysed during the current study are available from the corresponding author on reasonable request.

References

Sommer A, Taylor HR, Ravilla TD, West S, Lietman TM, Keenan JD, Chiang MF, Robin AL, Mills RP. Challenges of ophthalmic care in the developing world. JAMA Ophthalmol. 2014;132(5):640–4. https://doi.org/10.1001/jamaophthalmol.2014.84

Ting D, Ho CS, Deshmukh R, Said DG, Dua HS. Infectious keratitis: an update on epidemiology, causative microorganisms, risk factors, and antimicrobial resistance. Eye (Lond). 2021;35(4):1084–101. https://doi.org/10.1038/s41433-020-01339-3

Nguyen T, Ong J, Masalkhi M, Waisberg E, Zaman N, Sarker P, Aman S, Lin H, Luo M, Ambrosio R, et al. Artificial intelligence in corneal diseases: A narrative review. Cont Lens Anterior Eye. 2024;47(6):102284. https://doi.org/10.1016/j.clae.2024.102284

Cabrera-Aguas M, Watson SL. Updates in diagnostic imaging for infectious keratitis: A review. Diagnostics (Basel). 2023;13(21):3358. https://doi.org/10.3390/diagnostics13213358

Hu S, Sun Y, Li J, Xu P, Xu M, Zhou Y, Wang Y, Wang S, Ye J. Automatic diagnosis of infectious keratitis based on Slit lamp images analysis. J Pers Med. 2023;13(3):519. https://doi.org/10.3390/jpm13030519

Li Z, Jiang J, Chen K, Chen Q, Zheng Q, Liu X, Weng H, Wu S, Chen W. Preventing corneal blindness caused by keratitis using artificial intelligence. Nat Commun. 2021;12(1):3738. https://doi.org/10.1038/s41467-021-24116-6

Wu J, Yuan Z, Fang Z, Huang Z, Xu Y, Xie W, Wu F, Yao YF. A knowledge-enhanced transform-based multimodal classifier for microbial keratitis identification. Sci Rep. 2023;13(1):9003. https://doi.org/10.1038/s41598-023-36024-4

Xu Y, Kong M, Xie W, Duan R, Fang Z, Lin Y, Zhu Q, Tang S, Wu F, Yao Y. Deep sequential feature learning in clinical image classification of infectious keratitis. Engineering. 2021;7(07):245–63. https://doi.org/10.1016/j.eng.2020.04.012

Pandey PU, Ballios BG, Christakis PG, Kaplan AJ, Mathew DJ, Ong TS, Wan MJ, Micieli JA, Wong J. Ensemble of deep convolutional neural networks is more accurate and reliable than board-certified ophthalmologists at detecting multiple diseases in retinal fundus photographs. Br J Ophthalmol. 2024;108(3):417–23. https://doi.org/10.1136/bjo-2022-322183

Ong ZZ, Sadek Y, Qureshi R, Liu SH, Li T, Liu X, Takwoingi Y, Sounderajah V, Ashrafian H, Ting D, et al. Diagnostic performance of deep learning for infectious keratitis: a systematic review and meta-analysis. EClinicalMedicine. 2024;77:102887. https://doi.org/10.1016/j.eclinm.2024.102887

Gurnani B, Kaur K, Lalgudi VG, Kundu G, Mimouni M, Liu H, Jhanji V, Prakash G, Roy AS, Shetty R, et al. Role of artificial intelligence, machine learning and deep learning models in corneal disorders - A narrative review. J Fr Ophtalmol. 2024;47(7):104242. https://doi.org/10.1016/j.jfo.2024.104242

Li Z, Xie H, Wang Z, Li D, Chen K, Zong X, Qiang W, Wen F, Deng Z, Chen L, et al. Deep learning for multi-type infectious keratitis diagnosis: A nationwide, cross-sectional, multicenter study. NPJ Digit Med. 2024;7(1):181. https://doi.org/10.1038/s41746-024-01174-w

Tabuchi H, Engelmann J, Maeda F, Nishikawa R, Nagasawa T, Yamauchi T, Tanabe M, Akada M, Kihara K, Nakae Y, et al. Using artificial intelligence to improve human performance: efficient retinal disease detection training with synthetic images. Br J Ophthalmol. 2024;108(10):1430–5. https://doi.org/10.1136/bjo-2023-324923

Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A, Bengio Y. Generative adversarial networks. Commun ACM. 2020;63(11):139–44. https://doi.org/10.1145/3422622

Song J, Meng C, Ermon S. Denoising diffusion implicit models. ArXiv. 2020. https://doi.org/10.48550/arXiv.2010.02502. 2010.02502.

Rombach R, Blattmann A, Lorenz D, Esser P, Ommer B. High-Resolution image synthesis with latent diffusion models. ArXiv. 2021. https://doi.org/10.48550/arXiv.2112.10752. 2112.10752.

Lim JJ, Veasuvalingam B. Does online case-based learning foster clinical reasoning skills? A mixed-methods study. Future Healthc J. 2025;12(1):100210. https://doi.org/10.1016/j.fhj.2024.100210

Wang Y, Peng Y, Huang Y. The effect of typical case discussion and scenario simulation on the critical thinking of midwifery students: evidence from China. BMC Med Educ. 2024;24(1):340. https://doi.org/10.1186/s12909-024-05127-5

Choudhry NK, Fletcher RH, Soumerai SB. Systematic review: the relationship between clinical experience and quality of health care. Ann Intern Med. 2005;142(4):260–73. https://doi.org/10.7326/0003-4819-142-4-200502150-00008

Dash NR, Guraya SY, Al BM, Abdalla ME, Yusoff M, Al-Qahtani MF, van Mook W, Shafi MS, Almaramhy HH, Mukhtar W. Preferred teaching styles of medical faculty: an international multi-center study. BMC Med Educ. 2020;20(1):480. https://doi.org/10.1186/s12909-020-02358-0

Johnson MW, Gheihman G, Thomas H, Schiff G, Olson A, Begin AS. The impact of clinical uncertainty in the graduate medical education (GME) learning environment: A mixed-methods study. Med Teach. 2022;44(10):1100–8. https://doi.org/10.1080/0142159X.2022.2058383

Sonmez SC, Sevgi M, Antaki F, Huemer J, Keane PA. Generative artificial intelligence in ophthalmology: current innovations, future applications and challenges. Br J Ophthalmol. 2024;108(10):1335–40. https://doi.org/10.1136/bjo-2024-325458

Xie X, Jiachu D, Liu C, Xie M, Guo J, Cai K, Li X, Mi W, Ye H, Luo L, et al. Generating synthesized fluorescein angiography images from color fundus images by generative adversarial networks for macular edema assessment. Transl Vis Sci Technol. 2024;13(9):26. https://doi.org/10.1167/tvst.13.9.26

Albert FA, Seidu AA, Mason HM, Anderson E, Alele FO, Heggarty P, Hollins A, Gupta TS, Hays RB, McArthur L, et al. A systematic review of medical practitioners’ retention and application of basic sciences to clinical practice. BMC Med Educ. 2024;24(1):997. https://doi.org/10.1186/s12909-024-05952-8