ACRT Journal

Discussion

Collectively, these findings demonstrate that ChatGPT significantly improves the readability of breast imaging reports in both English and Spanish without sacrificing clinical accuracy. Improvements were consistent across all assessed readability metrics and across a diverse set of report types, including screening mammograms, diagnostic studies, ultrasounds, and reports involving biopsy. In English, ChatGPT reduced the reading level by approximately two grade levels on average (from FKRL 11.6 to 9.6), shifting many reports from a “difficult” readability range into a more “standard” or “relatively difficult” range, closer to public health recommendations for patient materials. The distribution histogram confirms that these shifts were not isolated to a few outliers but rather reflected a widespread pattern across the sample. In Spanish, the effect was even more striking. Based on the Fernández-Huerta and Szigriszt-Pazos scales, the vast majority of translated reports achieved “very easy” or “easy” readability (scores ≥ 80), with no translations falling into the “very difficult” or “difficult” ranges. This resulted in a noticeable shift toward lower-grade equivalents compared to the original English reports. Importantly, despite these simplifications, all 50 translated outputs retained full clinical accuracy, as verified by the bilingual breast radiologist.

Together, these findings suggest that ChatGPT can be a powerful tool to bridge the communication gap in breast imaging, particularly for Spanish-speaking populations who may face compounded barriers due to both language and literacy challenges. By significantly lowering readability scores and maintaining medical accuracy, ChatGPT outputs may help patients better understand their imaging results, potentially reducing anxiety, improving informed decision-making, and supporting shared decision-making conversations with healthcare providers [26].

Several recent studies have explored the use of LLMs to improve patient understanding of radiology reports, primarily through automated summarization and rewording strategies. In 2016, Hassanpour and Langlotz pioneered this area by developing natural language processing methods to extract key findings from radiology reports for structured summarization [27]. Furthermore, Van Veen et al. introduced RadAdapt, an LLM pipeline using lightweight domain adaptation to fine-tune models specifically on radiology reports [28]. Their model produced concise impressions that aligned with expert radiologist assessments, demonstrating both fluency and factual coherence in summarization tasks. Building on this trajectory, in 2025, Yang et al. employed reinforcement learning from AI feedback to align LLM-generated chest CT impressions with radiologist-preferred summaries, achieving ~78% agreement and measurable gains in precision, recall, and F1 scores over unaligned models [29].

Recent work has also begun to examine LLM performance in producing patient-directed summaries. In 2024, Jeblick et al. evaluated ChatGPT for rewriting radiology reports into plain language and found substantial improvements in readability while maintaining high factual accuracy—though some minor omissions were observed in less clinically significant details [12]. Kuckelman et al. examined ChatGPT-4 in the context of musculoskeletal radiology and demonstrated that the model could consistently produce simplified summaries that retained key diagnostic content, indicating its potential utility in patient communication workflows [30]. Similarly, Park et al. developed a generativeAI pipeline specifically tailored for creating patient-centered radiology reports [31]. Their study reported notable improvements in patients’ perceived clarity and ease of understanding, without compromising clinical accuracy or diagnostic integrity. These findings collectively suggest that LLMs may offer scalable solutions for improving the accessibility of radiologic information. However, despite these encouraging advances, most prior research has focused exclusively on English-language reports and clinicianfacing summarization tasks, rather than addressing the dual challenges of improving lay readability and meeting the needs of multilingual patient populations. Further work here is indicated.

While this study demonstrates the potential of ChatGPT to improve accessibility of breast imaging reports through simplified Spanish translations, several limitations merit discussion. First, there is no single readability metric validated for both English and Spanish text, limiting true statistical comparison of English and Spanish readability indices. Second, the dataset consisted of de-identified mammography and ultrasound reports, which may not represent the full range of breast imaging language encountered in diverse clinical settings. The selection may still be biased toward reports with ambiguous or complex wording, given the nature of patientgenerated content, and may not capture the variability present in institutional documentation styles. Third, the readability metrics used, Flesch-Kincaid Reading Level (FKRL) and FernándezHuerta, offer useful benchmarks but are inherently limited in their ability to assess true patient comprehension. These indices are based on surface-level textual features like sentence and word length, rather than semantic clarity, numeracy, or cultural relevance. Moreover, while both are well-established, they were developed for general prose and may not fully capture the complexities of bilingual or health-specific communication. Fourth, the study assessed translation quality and clinical accuracy based on expert review by a single bilingual, fellowship-trained academic breast radiologist. While this ensured a high level of clinical consistency, it limits the generalizability of our findings. Broader validation by panels including native Spanish-speaking patients, patient advocates, and multidisciplinary clinicians will be important to assess real-world utility and trust. Fifth, we evaluated only one LLM, ChatGPT-4o, at a single time point. As LLMs continue to evolve rapidly, future research should compare performance across different models (e.g., Gemini, Claude, Copilot) and evaluate the impact of newer model versions, system prompts, and fine-tuning strategies tailored to radiology or patient education.

Conclusion

This study demonstrates that ChatGPT can effectively simplify breast imaging reports in both English and Spanish, substantially improving readability while preserving clinical accuracy. By reducing the average reading level by approximately two grades in English and achieving “easy” or “very easy” readability in Spanish, ChatGPT addresses a critical communication gap for patients with varying health literacy and language needs. These findings highlight the potential of LLMs to enhance patient engagement and comprehension, particularly in breast imaging, where anxiety and misunderstanding of reports are common. As generative AI technologies continue to advance, integrating them into patient-centered communication strategies, while ensuring expert oversight, may help bridge linguistic and literacy barriers, ultimately improving informed decision-making and trust in radiologic care.

Acknowledgements None.

Ethical Considerations

This study was deemed exempt from IRB review by the IRB. Informed consent was waived because this study was determined not to be research involving human subjects by the IRB.

Funding

None.

Data Availability

The original contributions presented in this study are included in the article/supplementary material. Further inquiries can be directed to the corresponding author.

Conflicts of Interest

The authors declare no conflict of interest.

References

1. Alarifi M, Patrick T, Jabour A, Wu M, Luo J. (2021). Understanding patient needs and gaps in radiology reports through online discussion forum analysis. Insights Imaging. 12: 50.
2. Brett J, Bankhead C, Henderson B, Watson E, Austoker J. (2005). The psychological impact of mammographic screening. A systematic review. Psychooncology. 14: 917-38.
3. Pollock JR, Petty SAB, Schmitz JJ, Varner J, Metcalfe AM. (2024). Patient Access of Their Radiology Reports Before and After Implementation of 21st Century Cures Act Information-Blocking Provisions at a Large Multicampus Health System. AJR Am J Roentgenol. 222: e2330343.
4. McKee J. (2025). Radiology Adjusts to The 21st Century Cures Act.
5. Li H, Moon JT, Iyer D. (2023). Decoding radiology reports: Potential application of OpenAI ChatGPT to enhance patient understanding of diagnostic reports. Clin Imaging. 101: 137-141.
6. Rogers C, Willis S, Gillard S, Chudleigh J. (2023). Patient experience of imaging reports: A systematic literature review. Ultrasound. 31: 164- 175.
7. Munawar K, Prabhu V. (2021). Radiology on Reddit: A Content Analysis and Opportunity for Radiologist Engagement and Education. Curr Probl Diagn Radiol. 50: 362-368.
8. Dietrich S, Hernandez E. (2019). Nearly 68 Million People Spoke a
9. Villa Camacho JC, Pena MA, Flores EJ. (2021). Addressing Linguistic Barriers to Care: Evaluation of Breast Cancer Online Patient Educational Materials for Spanish-Speaking Patients. J Am Coll Radiol. 18: 919-926.
10. Choudhery S, Xi Y, Chen H. (2020). Readability and Quality of Online Patient Education Material on Websites of Breast Imaging Centers. J Am Coll Radiol. 17: 1245-1251.
11. AlKhalili R, Shukla PA, Patel RH, Sanghvi S, Hubbi B. (2015). Readability assessment of internet-based patient education materials related to mammography for breast cancer screening. Acad Radiol. 22: 290-295.
12. Jeblick K, Schachtner B, Dexl J. (2024). ChatGPT makes medicine easy to swallow: an exploratory case study on simplified radiology reports. Eur Radiol. 34: 2817-2825.
13. Tepe M, Emekli E. (2024). Assessing the Responses of Large Language Models (ChatGPT-4, Gemini, and Microsoft Copilot) to Frequently Asked Questions in Breast Imaging: A Study on Readability and Accuracy. Cureus. 16: e59960.
14. Chung EM, Zhang SC, Nguyen AT, Atkins KM, Sandler HM. (2023). Feasibility and acceptability of ChatGPT generated radiology report summaries for cancer patients. Digit Health. 9: 20552076231221620.
15. Gibson D, Jackson S, Shanmugasundaram R. (2024). Evaluating the Efficacy of ChatGPT as a Patient Education Tool in Prostate Cancer: Multimetric Assessment. J Med Internet Res. 26: e55939.
16. Eppler MB, Ganjavi C, Knudsen JE. (2023). Bridging the Gap Between Urological Research and Patient Understanding: The Role of Large Language Models in Automated Generation of Layperson’s Summaries. Urol Pract. 10: 436-443.
17. Butler JJ, Harrington MC, Tong Y. (2024). From jargon to clarity: Improving the readability of foot and ankle radiology reports with an artificial intelligence large language model. Foot Ankle Surg. 30: 331337.
18. Temperley HC, O’Sullivan NJ, Mac Curtain BM. (2024). Current applications and future potential of ChatGPT in radiology: A systematic review. J Med Imaging Radiat Oncol. 68: 257-264.
19. Sarangi PK, Lumbani A, Swarup MS. (2023). Assessing ChatGPT’s Proficiency in Simplifying Radiological Reports for Healthcare Professionals and Patients. Cureus. 15: e50881.
20. Rogasch JMM, Metzger G, Preisler M. (2023). ChatGPT: Can You Prepare My Patients for [J Nucl Med. 64: 1876-1879.
21. Cherla DV, Sanghvi S, Choudhry OJ, Liu JK, Eloy JA. (2012). Readability assessment of Internet-based patient education materials related to endoscopic sinus surgery. Laryngoscope. 122: 1649-1654.
22. Fajardo-Delgado D, Rodriguez-Coayahuitl L, Sánchez-Cervantes MG. (2025). Readability Formulas for Elementary School Texts in Mexican Spanish Applied Sciences. 15: 7259.
23. O’Sullivan L, Sukumar P, Crowley R, McAuliffe E, Doran P. (2020). Readability and understandability of clinical research patient information leaflets and consent forms in Ireland and the UK: a retrospective quantitative analysis. BMJ Open. 10: e037994.
24. Nash E, Bickerstaff M, Chetwynd AJ, Hawcutt DB, Oni L. (2023). The readability of parent information leaflets in paediatric studies. Pediatr Res. 94: 1166-1171.
25. Graham C, Reynard JM, Turney BW. (2015). Consent information leaflets - readable or unreadable? J Clin Urol. 8: 177-182.
26. van der Mee FAM, Ottenheijm RPG, Gentry EGS. (2025). The impact of different radiology report formats on patient information processing: a systematic review. Eur Radiol. 35: 2644-2657.
27. Hassanpour S, Langlotz CP. (2016). Information extraction from multiinstitutional radiology reports. Artif Intell Med. 66: 29-39.
28. Van VD, Van UC, Attias M. (2023). RadAdapt: Radiology Report Summarization via Lightweight Domain Adaptation of Large Language Models. presented at: In Proceedings of the 22nd Workshop on Biomedical Natural Language Processing and BioNLP Shared Tasks; Toronto, Canada.
29. Yang L, Zhou Y, Qi J. (2025). Aligning large language models with radiologists by reinforcement learning from AI feedback for chest CT reports. Eur J Radiol. 184: 111984.
30. Kuckelman IJ, Wetley K, Yi PH, Ross AB. (2024). Translating musculoskeletal radiology reports into patient-friendly summaries using ChatGPT-4. Skeletal Radiol. 53: 1621-1624.
31. Park J, Oh K, Han K, Lee YH. (2024). Patient-centered radiology reports with generative artificial intelligence: adding value to radiology reporting. Sci Rep. 14: 13218.