{"paper_id":"26af4bd0-bec2-4e6a-bec4-40cd2e9a80a4","body_text":"Clinical performance tradeoffs of ChatGPT-5.2 Thinking (OpenAI) compared with radiologist interpretation in biopsy-referred mammography: cancer detection, false positives, and laterality | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Clinical performance tradeoffs of ChatGPT-5.2 Thinking (OpenAI) compared with radiologist interpretation in biopsy-referred mammography: cancer detection, false positives, and laterality Mohammad Alarifi, Abdulrahman Jabour, Ahmad Abanomy, Haitham Alahmad, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8701935/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Purpose To compare ChatGPT-5.2 Thinking (OpenAI) with practicing radiologists for the clinically relevant, examination-level task of classifying biopsy-proven malignancy in a biopsy-referred mammography test set, and to assess performance by radiologic feature type and laterality Methods In this multicenter retrospective study across several cities in Saudi Arabia, screening mammograms from an initial cohort of 1,225 women were linked to pathology to create a biopsy-anchored reference standard. Board-certified breast radiologists, blinded to pathology and model outputs, provided study-level BI-RADS® assessments. ChatGPT-5.2 Thinking received de-identified bilateral CC and MLO views with a fixed BI-RADS–based prompt and produced an ordinal BI-RADS category (0–5) and suspected laterality. The analytic test set included 100 examinations that proceeded to biopsy (61 biopsy-confirmed cancers and 39 biopsy-negative controls). Primary outcomes were case-level sensitivity, specificity, and accuracy; secondary outcomes included laterality performance and feature-level patterns. All analyses were executed in Python and R on secured institutional workstations. Results ChatGPT-5.2 demonstrated higher sensitivity than radiologists (95.1% vs 82.0%) but lower specificity (10.3% vs 56.4%), yielding lower overall accuracy (62.0% vs 72.0%). Feature-wise, AI showed highest sensitivity with dense parenchymal patterns and highest specificity for architectural distortion, tended to overcall mass-like findings, and performed weakest for microcalcifications. Laterality accuracy was 60.7%. Conclusion In this biopsy-referred, pathology-anchored evaluation, ChatGPT-5.2 Thinking showed higher sensitivity but substantially lower specificity than radiologists, supporting its potential role as a concurrent aid/triage signal rather than a stand-alone reader, pending prospective validation. Artificial intelligence Mammography Breast cancer BI-RADS Diagnostic accuracy Reader study Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Breast cancer remains a leading cause of cancer-related morbidity and mortality among women worldwide and continues to place a substantial clinical and economic burden on health systems [1, 2]. Screening mammography has been central to early detection programs and has contributed to stage migration and improvements in survival through earlier diagnosis and timely referral to definitive therapy [3, 4]. Despite these gains, the performance of screening remains variable across programs and settings. Reader sensitivity and specificity depend on multiple factors that include breast density, lesion conspicuity, workload, double-reading policies, training, and access to prior images for comparison [5–8]. False negatives may delay diagnosis while false positives may trigger unnecessary callbacks, added imaging, anxiety, and biopsies that do not yield cancer [9–11]. These realities have motivated sustained efforts to enhance accuracy while preserving efficiency in high-volume screening environments [12, 13]. Artificial intelligence-based systems for mammography have emerged as a promising approach to support detection and decision making [14, 15]. Deep-learning methods trained on large datasets can learn visual patterns associated with masses, microcalcifications, architectural distortion, and asymmetries, and then produce image-level or case-level predictions that can be integrated into the reading workflow [16, 17]. Several studies report that AI can match or approach single-reader performance and can assist readers through triage, prioritization, or second reader-style support [18–21]. At the same time, the literature also documents failure modes that include difficulties with subtle calcifications, architectural distortion, and laterality or localization errors, as well as performance variability across vendors, acquisition protocols, and clinical settings [22–25]. There is an ongoing need for rigorous evaluation against robust reference standards and for transparent reporting of metrics that matter to patients and clinicians [26, 27]. In parallel, patients increasingly view their reports directly via portals, and a recent systematic review shows sustained patient demand for access alongside comprehension- and anxiety-related challenges; therefore, readability and patient-centred communication should be explicit design criteria when integrating AI outputs into screening workflows [28]. In this context, patient-facing AI interpretations could function as a structured ‘second look’ explanation that helps patients formulate questions and participate in shared decision-making during clinical consultations. However, such use should be positioned as decision support that complements rather than replaces radiologist assessment and clinician-guided management Biopsy confirmation offers a decisive reference to evaluate diagnostic systems because it anchors analyses to ground truth rather than to follow-up imaging alone [29]. Cohorts built on biopsy-confirmed cancers allow unbiased estimation of sensitivity and enable stratified analyses by lesion type, size, and breast density [29, 30]. When paired with an appropriate sample of biopsy-negative controls or verified benign outcomes, these cohorts also enable meaningful estimation of specificity and accuracy [30]. In addition, expert review of discordant cases can reveal systematic error patterns that inform future model development, calibration, and deployment strategies [31]. The present study evaluates ChatGPT-5.2 Thinking (OpenAI) as a stand-alone model for the clinically defined imaging task of examination-level malignancy classification using bilateral CC and MLO screening mammography views in a biopsy-referred cohort with pathology reference. We hypothesized that the model would achieve sensitivity comparable to radiologists for biopsy-proven malignancy but would demonstrate lower specificity due to benign overcalling in a biopsy-enriched test set. We additionally assessed laterality performance and feature-type error patterns (mass, architectural distortion, microcalcifications, and parenchymal density) to identify clinically actionable failure modes [32]. 2. Methods 2.1. Design and setting Retrospective, multicenter diagnostic-accuracy study conducted across several cities in Saudi Arabia within participating breast imaging centers. An upstream screening population of 1,225 women was accrued from multiple health systems during the study period. From this screened cohort, we constructed a biopsy-referred test set by including examinations that proceeded to image-guided or surgical biopsy based on clinical assessment. The analytic test set comprised 100 biopsy-referred mammography examinations (four standard 2D views per exam; 400 images total), including 61 biopsy-confirmed malignancies and 39 biopsy-negative controls, linked deterministically to each site’s pathology database to establish a biopsy-anchored reference standard. This design was selected to ensure definitive ground truth for malignancy status and to enable direct estimation of sensitivity and specificity against pathology. However, because the cohort is enriched for biopsy-referred cases, it is not representative of a general screening population and may yield inflated sensitivity and depressed specificity relative to true screening prevalence; we therefore interpret performance as applicable to a biopsy-referred triage / concurrent-aid scenario rather than population screening. Accordingly, we quantified case-level sensitivity, specificity, and accuracy for malignant vs benign biopsy outcome (primary endpoint), and evaluated laterality performance and feature-type patterns (secondary endpoints). 2.2. Imaging and readers For each case, bilateral craniocaudal (CC) and mediolateral oblique (MLO) views were exported as DICOM files and de-identified; each case comprised four images. Board-certified breast radiologists were blinded to pathology and to ChatGPT-5.2 Thinking outputs. Radiologists provided study-level BI-RADS® assessment categories (0–5) and documented the suspected laterality (right vs left) of the abnormality prompting the assessment. Radiologist BI-RADS categories were used to generate binary calls at pre-specified thresholds for comparison with pathology and to classify true-positive, false-positive, true-negative, and false-negative outcomes. Laterality was used to define side-specific correctness among positive cases. ChatGPT-5.2 Thinking was evaluated in stand-alone mode using the same de-identified mammography views provided case-by-case (four images per examination) with a fixed, pre-specified BI-RADS–guided prompt (Fig. 1 ). The model generated an ordinal BI-RADS category (0–5) and suspected laterality for each case. ChatGPT-5.2 Thinking was not used to create or alter images, define ground truth, or compute statistical metrics; all performance calculations were performed by the study team using pathology as the reference standard. 2.3. Artificial intelligence and prompting ChatGPT-5.2 Thinking (OpenAI) was the only AI system evaluated in this study. The model was used in stand-alone mode to interpret de-identified bilateral CC and MLO mammography views. A fixed, pre-specified prompt instructed the model to produce: (i) an ordinal BI-RADS® assessment category (0–5) using BI-RADS-consistent terminology, (ii) a binary malignancy classification derived from a pre-defined BI-RADS threshold, and (iii) suspected laterality (right vs left) when an abnormality was identified. The prompt was locked prior to evaluation and applied uniformly across cases (Fig. 1 ). ChatGPT-5.2 Thinking was not used to generate ground-truth labels, alter images, select cases, or compute performance statistics; all metrics were calculated by the study team using pathology as the reference standard. 2.4. Statistical analysis, software, and reproducibility Categorical variables were summarized as counts and percentages, and continuous variables as means with standard deviations or medians with interquartile ranges, according to distributional shape. Primary case-level performance metrics were sensitivity, specificity, and overall accuracy; laterality accuracy (right vs left breast) was analyzed as a secondary endpoint. Exact 95% confidence intervals were calculated for proportions using the Clopper Pearson method, and paired proportions were compared with McNemar tests. Threshold-based summaries (e.g., BI-RADS ≥ 3/≥4/≥5) are reported with corresponding true-positive and false-positive counts. Missing data were handled with complete-case analysis. All hypothesis tests were two-sided with α = 0.05. Pre-specified subgroup analyses included lesion “sign” (mass, microcalcifications, architectural distortion, parenchymal density) and breast density categories when sample size permitted stable estimates. All analyses were executed in Python and R on secured institutional workstations. Reproducible pipelines were implemented as version-controlled scripts with fixed random seeds, where applicable; summary tables and figures were generated directly from analysis outputs. Parameter settings, analysis logs, and software versions were archived alongside the study protocol to facilitate replication. Institutional oversight was provided by the University of Wisconsin–Milwaukee IRB (IRB #20.230), with a waiver of informed consent for the use of de-identified, minimal-risk data. Data availability: De-identified imaging data cannot be publicly shared due to patient-privacy restrictions and institutional policy. Analysis code resides on an access-controlled, encrypted institutional workstation under the corresponding author’s account and can be provided to qualified researchers upon reasonable request, subject to IRB oversight and a data-use agreement. 3. Results 3.1. Diagnostic Performance Comparison between Radiologists and AI Figure 2 presents a comparative analysis of radiologist and AI diagnostic performance metrics, using biopsy results as the gold standard. The AI model demonstrated a higher sensitivity (95.08%) than radiologists (81.97%), indicating a stronger ability to identify biopsy-confirmed malignant cases. However, this improvement came at the expense of specificity, which dropped sharply to 10.26% compared to 56.41% for radiologists. Consequently, radiologists achieved a higher overall accuracy (72.00%) compared to AI (62.00%). In absolute terms, ChatGPT-5.2 correctly identified 58 true-positive cases but also generated 35 false-positive predictions, while radiologists recorded 50 true positives and 17 false positives. The AI missed only three true malignancies (FN = 3), whereas radiologists missed eleven (FN = 11). Although AI demonstrated better sensitivity, expert image review revealed that many of its false-positive detections corresponded to benign structures or peripheral artifacts, indicating misclassification rather than true lesion recognition. 3.2. AI Performance by Radiologic Feature Type Figure 3 summarizes the AI model’s diagnostic metrics across four radiologic lesion categories: Mass, Architectural Distortion (AS), Microcalcifications (MC), and Parenchymal Density (PD). Among all lesion types, parenchymal density (PD) achieved the highest sensitivity (0.818), suggesting that AI performed best when detecting diffuse density changes. Mass lesions also demonstrated relatively high sensitivity (0.714) but suffered from extremely low specificity (0.125), reflecting a tendency of AI to misclassify normal dense areas as suspicious masses. Architectural distortion (AS) showed a balanced performance, with sensitivity of 0.455 and high specificity of 0.84 the most favorable trade-off among all categories. In contrast, microcalcifications (MC) were the weakest feature type for AI, with sensitivity of only 0.25 and precision of 0.211, indicating poor recognition of small-scale calcified foci. Overall, the pattern suggests that AI performs relatively well in identifying large-area abnormalities (PD, mass) but lacks reliability in localizing fine-grained structures (MC, AS). This variability across lesion types aligns with radiologists’ qualitative observations that AI often highlights regions adjacent to, rather than directly over, true lesions. 3.3. AI Performance in Breast-Side Localization Figure 4 displays the AI’s accuracy in correctly identifying the laterality (right vs. left breast) of detected abnormalities. The model correctly localized the disease in 37 cases (“True side”) and failed in 24 cases (“False side”). This yields an overall laterality accuracy of 60.66%, reflecting moderate spatial awareness but insufficient precision for clinical reliability. In many incorrect cases, AI successfully detected the presence of a lesion but marked it on the contralateral breast or outside the actual parenchymal area. Such errors underscore the model’s limitations in spatial alignment and contextual awareness, particularly in cases with overlapping tissue density or bilateral findings. 4. Discussion In this biopsy-referred, pathology-anchored test set, ChatGPT-5.2 Thinking demonstrated higher sensitivity but markedly lower specificity than radiologists, resulting in higher false-positive prompting. Clinically, this pattern is consistent with a tool that may help reduce missed cancers if used as a concurrent aid (e.g., prompting a “second look” on cases otherwise read as benign), but it may also increase additional workup and patient anxiety due to benign overcalling. Beyond workflow impact, accessible AI outputs may also support patient engagement by helping patients understand the reason for further assessment and arrive to visits prepared for a more focused discussion. Any patient-facing use should remain tightly coupled to clinician review to avoid misinterpretation and unintended delays in care. This tradeoff is consistent with prior evaluations where deep learning systems increase case finding but introduce benign overcalls under real world mix of screening and diagnostic studies [1, 2, 4, 6, 11, 17, 22, 24]. Therefore, the most appropriate near-term clinical framing is triage/prioritization or concurrent-aid support, not replacement of radiologist interpretation [5, 10, 15, 20]. Feature level behavior clarifies where artificial intelligence helped and where it struggled. Parenchymal density produced the highest sensitivity, which aligns with studies showing that textural and global parenchymal patterns are well captured by convolutional and transformer-based encoders trained at image scale [7, 13, 18]. Architectural distortion achieved the best specificity but only modest sensitivity, a conservative posture that mirrors prior reports in which models avoid distortion overcalls yet miss subtle tethering or spiculations in dense tissue [9, 14, 18]. Mass showed moderate sensitivity but very low specificity, consistent with the known risk of mistaking overlapping tissue and benign masses for malignancy, especially in heterogeneously dense breasts [3, 5, 19, 29]. Microcalcifications remained the weakest category, echoing literature that small, sparse, and high frequency signals are sensitive to pixel resolution, preprocessing, and training set curation [8, 12, 23, 27]. Together these patterns suggest practical parameterization: use conservative thresholds when calcifications or mass like appearances are suspected, while allowing artificial intelligence more freedom to surface diffuse parenchymal abnormalities [10, 13, 18]. Laterality accuracy was 60.7%, which represents a meaningful constraint for workflow. Wrong side assignment can lead to additional views, unnecessary callbacks, and avoidable anxiety, even when the case is correctly classified as positive [3, 16, 26]. Multiple groups have improved side awareness by enforcing cross view consistency checks and by incorporating quadrant aware attention maps or side conditioned decoders [7, 17, 31]. Our results support adoption of such side aware constraints before routine deployment. Breast density influenced both artificial intelligence and reader performance. Sensitivity decreased in heterogeneously and extremely dense categories, consistent with lesion masking from x ray attenuation and reduced conspicuity [1, 4, 27]. Studies that integrate prior examinations, tomosynthesis, or targeted ultrasound have documented partial recovery of sensitivity, suggesting that artificial intelligence systems which fuse priors or cross modality context may mitigate density related loss [6, 17, 29]. Until such fusion is widely available, cautious interpretation in dense breasts and targeted secondary imaging remain necessary. Comparison with published benchmarks indicates that our radiologist accuracy aligns with contemporary single reader performance, while the artificial intelligence profile falls within the range reported for research grade systems evaluated outside the original development distribution [2, 11, 17, 24]. Importantly, several multi center studies have shown that artificial intelligence specificity is most unstable when prevalence, vendor mix, and positioning differ from training data [9, 14, 30, 33, 34]. This reinforces the value of local calibration prior to adoption. As an operational model, a concurrent-aid approach elevates higher-risk studies in the worklist and cues focused assessment yet maintains independent judgment [5, 15, 20, 35, 36]. To reduce benign overcalls, some programs apply decision curves that trade small losses in sensitivity for meaningful reductions in unnecessary recalls [2, 24, 37]. Artificial intelligence as an independent second reader may substitute for double reading in resource constrained settings only after specificity and laterality accuracy are improved and clear escalation rules for discordant cases are validated prospectively [4, 22, 30]. User interface design remains critical to avoid automation bias and to protect the benefits of independent reads [10, 19]. Safety and governance require noting that, while laboratory anchoring reduces verification bias for malignancy, benign ground truth still partly depends on imaging stability, which can skew specificity estimates [1, 35, 38]. Prospective quality assurance should include periodic review of false positives and false negatives with feedback loops to both model and readers, an approach associated with measurable gains in other imaging domains [18, 39, 40]. Transparent reporting with confidence intervals, calibration plots, and decision curves enables oversight bodies to judge tradeoffs that reflect local priorities [2, 29, 41]. Equity and generalizability require attention, as documented performance drift across vendors, compression paddles, and population subgroups can affect both sensitivity and specificity [9, 14, 30]. External validation across diverse sites with predefined acceptability margins and periodic recalibration after deployment is therefore essential [17, 30, 31]. Adherence to reporting standards and public documentation of thresholds and update cadence support reproducibility and fair comparison across systems [19, 29]. Implications for clinical practice: Given its higher sensitivity but lower specificity, artificial intelligence is best used to trigger second looks on studies otherwise read as negative, helping reduce missed cancers [5, 10]. However, because benign overcalls carry cost and anxiety, programs should track recall rate, positive predictive value of biopsy recommendation, and time to diagnostic resolution after adoption [2, 22, 24]. The feature type profile suggests that readers should be most skeptical of artificial intelligence prompts for mass in dense tissue and for subtle calcifications, while considering artificial intelligence prompts for diffuse parenchymal patterns as higher yield [8, 13, 18]. To boost specificity, augment training sets with common benign look-alikes including tissue overlap, fibroadenomas with calcifications, and postsurgical change while employing cost-sensitive losses that penalize false positives [8, 12, 19]. Second, improve calcification detection by increasing native pixel resolution, adding calcification specific augmentations, and using multi scale heads tuned for small object detection [8, 23, 27]. Third, harden laterality by enforcing side aware constraints and cross view agreement, and by rejecting outputs that conflict with view geometry [3, 16, 26, 31]. Fourth, explore fusion of priors and tomosynthesis to mitigate density effects and improve robustness across acquisition variability [6, 17, 29, 30]. 5. Conclusion Artificial intelligence approached radiologist level detection but had lower specificity and therefore lower overall accuracy. Performance differed by feature type, with strengths in PD, conservative high specificity for AD, overcalling in MAS, and the weakest performance in MC. Laterality accuracy was modest at sixty-point seven percent, highlighting a practical barrier to deployment. These findings support artificial intelligence as a concurrent aid or prioritization tool rather than a stand-alone reader, and motivate targeted refinement in specificity, microcalcification handling, and breast side assignment, followed by prospective and external validation. Abbreviations AI, artificial intelligence BI-RADS, Breast Imaging Reporting and Data System CC, craniocaudal MLO, mediolateral oblique Declarations The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Competing Interests The authors declare no relevant financial or non-financial interests related to this work. Ethics approval and consent to participate Ethics approval and consent to participate: The study was approved by the University of Wisconsin–Milwaukee Institutional Review Board (IRB #20.230), and was conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines and regulations; informed consent was waived by the IRB because the study used de-identified, minimal-risk retrospective data. Funding This work was supported by King Saud University. Mohammad Alarifi has received research support from King Saud University (KSU). Author Contribution Conceptualization: Alarifi, AlmanaaMethodology: Alarifi, Jabour, AbanomyValidation: Aloufi, AlahmadFormal analysis: Alarifi, Jabour, AbanomyInvestigation: Aloufi, Abanomy, Alahmad, Alenazi, AlmanaaResources: Alenazi, AlmanaaData curation: Aloufi, JabourVisualization: Alarifi, AlshediWriting – original draft: Alarifi, Aloufi, JabourWriting – review & editing: All authorsSupervision: Alarifi, AlmanaaProject administration: AlarifiAll authors read and approved the final manuscript. Data Availability De-identified imaging data cannot be publicly shared due to patient-privacy restrictions and institutional policy. Analysis code resides on an access-controlled, encrypted institutional workstation under the corresponding author’s account and can be provided to qualified researchers upon reasonable request, subject to IRB oversight and a data-use agreement. References Rodríguez-Ruiz A, Krupinski E, Mordang J-J, Schilling K, Heywang-Köbrunner SH, Sechopoulos I, et al (2019) Detection of breast cancer with mammography: effect of an artificial intelligence support system. Radiology 290(2):305–314 . Resch D, Lo Gullo R, Teuwen J, Semturs F, Hummel J, Resch A, et al (2024) AI-enhanced mammography with digital breast tomosynthesis for breast cancer detection: clinical value and comparison with human performance. Radiol Imaging Cancer 6(4):e230149 . van Winkel SL, Rodríguez-Ruiz A, Appelman L, Gubern-Mérida A, Karssemeijer N, Teuwen J, et al (2021) Impact of artificial intelligence support on accuracy and reading time in breast tomosynthesis image interpretation: a multi-reader multi-case study. Eur Radiol 31(11):8682–8691 . Park EK, Kwak S, Lee W, Choi JS, Kooi T, Kim E-K (2024) Impact of AI for digital breast tomosynthesis on breast cancer detection and interpretation time. Radiol Artif Intell 6(3):e230318 . Hadadi I, Rae W, Clarke J, McEntee M, Ekpo E (2021) Breast cancer detection: comparison of digital mammography and digital breast tomosynthesis across non-dense and dense breasts. Radiography 27(4):1027–1032 . Lauritzen AD, Rodríguez-Ruiz A, von Euler-Chelpin MC, Lynge E, Vejborg I, Nielsen M, et al (2022) An artificial intelligence–based mammography screening protocol for breast cancer: outcome and radiologist workload. Radiology 304(1):41–49 . Kim H-E, Kim HH, Han B-K, Kim KH, Han K, Nam H, et al (2020) Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multireader study. Lancet Digit Health 2(3):e138–e148 . Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Broeders M, Gennaro G, Clauser P, et al (2019) Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. J Natl Cancer Inst 111(9):916–922 . Choi WJ, An JK, Woo JJ, Kwak HY (2022) Comparison of diagnostic performance in mammography assessment: radiologist with reference to clinical information versus standalone artificial intelligence detection. Diagnostics (Basel) 13(1):117 . Lee SE, Hong H, Kim E-K (2024) Diagnostic performance with and without artificial intelligence assistance in real-world screening mammography. Eur J Radiol Open 12:100545 . Pacilè S, Lopez J, Chone P, Bertinotti T, Grouin JM, Fillard P (2020) Improving breast cancer detection accuracy of mammography with the concurrent use of an artificial intelligence tool. Radiol Artif Intell 2(6):e190208 . de Vries CF, Colosimo SJ, Staff RT, Dymiter JA, Yearsley J, Dinneen D, et al (2023) Impact of different mammography systems on artificial intelligence performance in breast cancer screening. Radiol Artif Intell 5(3):e220146 . Kim YS, Jang M-J, Lee SH, Kim S-Y, Ha SM, Kwon BR, et al (2022) Use of artificial intelligence for reducing unnecessary recalls at screening mammography: a simulation study. Korean J Radiol 23(12):1241 . Anderson AW, Marinovich ML, Houssami N, Lowry KP, Elmore JG, Buist DS, et al (2022) Independent external validation of artificial intelligence algorithms for automated interpretation of screening mammography: a systematic review. J Am Coll Radiol 19(2):259–273 . Mansour S, Sweed E, Gomaa MMM, Hussein SA, Abdallah E, Nada YM, et al (2024) Artificial intelligence as an initial reader for double reading in breast cancer screening: a prospective initial study of 32,822 mammograms of the Egyptian population. Egypt J Radiol Nucl Med 55(1):181 . Lee JH, Kim KH, Lee EH, Ahn JS, Ryu JK, Park YM, et al (2022) Improving the performance of radiologists using artificial intelligence-based detection support software for mammography: a multi-reader study. Korean J Radiol 23(5):505 . Nanaa M, Gupta VO, Hickman SE, Allajbeu I, Payne NR, Arponen O, et al (2024) Accuracy of an artificial intelligence system for interval breast cancer detection at screening mammography. Radiology 312(2):e232303 . Larsen M, Olstad CF, Lee CI, Hovda T, Hoff SR, Martiniussen MA, et al (2024) Performance of an artificial intelligence system for breast cancer detection on screening mammograms from BreastScreen Norway. Radiol Artif Intell 6(3):e230375 . Lång K, Hofvind S, Rodríguez-Ruiz A, Andersson I (2021) Can artificial intelligence reduce the interval cancer rate in mammography screening? Eur Radiol 31(8):5940–5947 . Larsen M, Aglen CF, Hoff SR, Lund-Hanssen H, Hofvind S (2022) Possible strategies for use of artificial intelligence in screen-reading of mammograms, based on retrospective data from 122,969 screening examinations. Eur Radiol 32(12):8238–8246 . Alarifi M (2025) Radiologists’ views on AI and the future of radiology: insights from a US national survey. Br J Radiol :tqaf222 . Zeng A, Houssami N, Noguchi N, Nickel B, Marinovich ML (2024) Frequency and characteristics of errors by artificial intelligence in reading screening mammography: a systematic review. Breast Cancer Res Treat 207(1):1–13 . AbuAbeileh EA, Abuabeileh Z, AbuAbeileh Z (2025) Screening mammography and artificial intelligence: a comprehensive systematic review. Cureus 17(2):e . Lång K, Dustler M, Dahlblom V, Åkesson A, Andersson I, Zackrisson S (2021) Identifying normal mammograms in a large screening population using artificial intelligence. Eur Radiol 31(3):1687–1692 . Alarifi M (2025) Appropriateness of thyroid nodule cancer risk assessment and management recommendations provided by large language models. J Imaging Inform Med :1–12 . Uematsu T, Nakashima K, Harada TL, Nasu H, Igarashi T (2023) Comparisons between artificial intelligence computer-aided detection synthesized mammograms and digital mammograms when used alone and in combination with tomosynthesis images in a virtual screening setting. Jpn J Radiol 41(1):63–70 . Lee SE, Yoon JH, Son N-H, Han K, Moon HJ (2024) Screening in patients with dense breasts: comparison of mammography, artificial intelligence, and supplementary ultrasound. AJR Am J Roentgenol 222(1):e2329655 . Alarifi M, Hughes MC, Jabour AM, Alashban Y, Vernon E (2024) Patient, referring physician, and radiologist opinions over time on providing patients access to radiology reports: a systematic review. J Am Coll Radiol 21(12):1862–1874 . Dembrower K, Crippa A, Colón E, Eklund M, Strand F (2023) Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. Lancet Digit Health 5(10):e703–e711 . Romero-Martín S, Elías-Cabot E, Raya-Povedano JL, Gubern-Mérida A, Rodríguez-Ruiz A, Álvarez-Benito M (2022) Stand-alone use of artificial intelligence for digital mammography and digital breast tomosynthesis screening: a retrospective evaluation. Radiology 302(3):535–542 . Watanabe AT, Lim V, Vu HX, Chim R, Weise E, Liu J, et al (2019) Improved cancer detection using artificial intelligence: a retrospective evaluation of missed cancers on mammography. J Digit Imaging 32(4):625–637 . Alarifi M (2025) Exploring radiologists’ use of AI chatbots for assistance in image interpretation: patterns of use and trust evaluation. J Imaging Inform Med :1–10 . Alabousi M, Wadera A, Al-Ghita MK, Al-Ghetaa RK, Salameh J-P, Pozdnyakov A, et al (2021) Performance of digital breast tomosynthesis, synthetic mammography, and digital mammography in breast cancer screening: a systematic review and meta-analysis. J Natl Cancer Inst 113(6):680–690 . Chikarmane SA, Cochon LR, Khorasani R, Sahu S, Giess CS (2021) Screening mammography performance metrics of 2D digital mammography versus digital breast tomosynthesis in women with a personal history of breast cancer. AJR Am J Roentgenol 217(3):587–594 . Dembrower K, Wåhlin E, Liu Y, Salim M, Smith K, Lindholm P, et al (2020) Effect of artificial intelligence-based triaging of breast cancer screening mammograms on cancer detection and radiologist workload: a retrospective simulation study. Lancet Digit Health 2(9):e468–e474 . Rodriguez-Ruiz A, Lång K, Gubern-Merida A, Teuwen J, Broeders M, Gennaro G, et al (2019) Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? a feasibility study. Eur Radiol 29(9):4825–4832 . Dang L-A, Chazard E, Poncelet E, Serb T, Rusu A, Pauwels X, et al (2022) Impact of artificial intelligence in breast cancer screening with mammography. Breast Cancer 29(6):967–977 . Sasaki M, Tozaki M, Rodríguez-Ruiz A, Yotsumoto D, Ichiki Y, Terawaki A, et al (2020) Artificial intelligence for breast cancer detection in mammography: experience of use of the ScreenPoint Medical Transpara system in 310 Japanese women. Breast Cancer 27(4):642–651 . Hadadi I, Rae W, Clarke J, McEntee M, Ekpo E (2022) Breast cancer detection across dense and non-dense breasts: markers of diagnostic confidence and efficacy. Acta Radiol Open 11(1):20584601211072279 . Kolb TM, Lichy J, Newhouse JH (2002) Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations. Radiology 225(1):165–175 . Yoon JH, Strand F, Baltzer PA, Conant EF, Gilbert FJ, Lehman CD, et al (2023) Standalone AI for breast cancer detection at screening digital mammography and digital breast tomosynthesis: a systematic review and meta-analysis. Radiology 307(5):e222639 . Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-8701935\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Research Article\",\"associatedPublications\":[],\"authors\":[{\"id\":585024676,\"identity\":\"f4cd5fe8-4644-466b-95e6-1505cfec9946\",\"order_by\":0,\"name\":\"Mohammad Alarifi\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuElEQVRIiWNgGAWjYFAC5gYGBgM5Hn4ojxgtjCAtxjySDaRpYTBmMDhArBbd9oONnysKDGSMb+SYbmCosE5sIKTF7Exis+QZAwMesxs5ZjcYzqQToeVAYoNkg8EfoJbcbTcY2w4ToeX8w+afDUBbjGeAtPwjRsuNxDZJkBYDCZCWBqK0PGyzBGmROPP+242EY+nGRDgs+fDNhj8G9vztaWk3PtRYyxLUggoSSFM+CkbBKBgFowAXAADHMkFATuZTggAAAABJRU5ErkJggg==\",\"orcid\":\"\",\"institution\":\"University of Wisconsin–Milwaukee\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Mohammad\",\"middleName\":\"\",\"lastName\":\"Alarifi\",\"suffix\":\"\"},{\"id\":585024677,\"identity\":\"c3342c7c-21ab-445c-be30-83ff59d9435f\",\"order_by\":1,\"name\":\"Abdulrahman Jabour\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Jazan University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Abdulrahman\",\"middleName\":\"\",\"lastName\":\"Jabour\",\"suffix\":\"\"},{\"id\":585024678,\"identity\":\"f9e946a7-f9ad-483d-98be-cd92c880604f\",\"order_by\":2,\"name\":\"Ahmad Abanomy\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Ahmad\",\"middleName\":\"\",\"lastName\":\"Abanomy\",\"suffix\":\"\"},{\"id\":585024682,\"identity\":\"aecd6eb1-6d0b-4e31-9ad8-a7a9fd25ff0b\",\"order_by\":3,\"name\":\"Haitham Alahmad\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Haitham\",\"middleName\":\"\",\"lastName\":\"Alahmad\",\"suffix\":\"\"},{\"id\":585024684,\"identity\":\"79e9a727-612b-4468-bba4-f48bc5d6b9c2\",\"order_by\":4,\"name\":\"Khaled Alenazi\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Khaled\",\"middleName\":\"\",\"lastName\":\"Alenazi\",\"suffix\":\"\"},{\"id\":585024690,\"identity\":\"3b941d1d-d044-49fc-a254-ab3db8efa4cf\",\"order_by\":5,\"name\":\"Alhanouf Alshedi\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Alhanouf\",\"middleName\":\"\",\"lastName\":\"Alshedi\",\"suffix\":\"\"},{\"id\":585024691,\"identity\":\"d4b62829-72cf-43cd-b564-effd4f141af0\",\"order_by\":6,\"name\":\"Mansour Almanaa\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mansour\",\"middleName\":\"\",\"lastName\":\"Almanaa\",\"suffix\":\"\"},{\"id\":585024694,\"identity\":\"ed7b7d8e-7be9-4372-918d-c3e171c276d4\",\"order_by\":7,\"name\":\"Areej Aloufi\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"King Saud University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Areej\",\"middleName\":\"\",\"lastName\":\"Aloufi\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2026-01-26 15:38:32\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-8701935/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-8701935/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":102180730,\"identity\":\"76c30dea-59b2-4eed-af80-62d8762e9bd1\",\"added_by\":\"auto\",\"created_at\":\"2026-02-09 07:13:42\",\"extension\":\"jpg\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":63069,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eLocked prompting workflow used for evaluation of ChatGPT-5.2 Thinking on de-identified bilateral CC and MLO views, producing BI-RADS (0–5) and suspected laterality..\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"1.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8701935/v1/f7cf2dd28ba1174f7131f86f.jpg\"},{\"id\":102180717,\"identity\":\"c75a7f08-69ef-4507-b7ae-a42c7a0b482c\",\"added_by\":\"auto\",\"created_at\":\"2026-02-09 07:13:37\",\"extension\":\"jpg\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":42730,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eDiagnostic performance tradeoffs of AI versus radiologists in a biopsy-referred cohort (sensitivity, specificity, and accuracy)\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"2.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8701935/v1/4e76c20848e12b17afa854b7.jpg\"},{\"id\":102180726,\"identity\":\"f6a6f860-a467-4779-899f-627102fdf4ea\",\"added_by\":\"auto\",\"created_at\":\"2026-02-09 07:13:41\",\"extension\":\"jpg\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":81224,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eAI performance stratified by radiologic feature type (architectural distortion, microcalcifications, and parenchymal density).\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"3.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8701935/v1/d23b45d4f351a794bc9348c8.jpg\"},{\"id\":102180731,\"identity\":\"621688e5-526c-4ded-ad01-b2bab8ee8476\",\"added_by\":\"auto\",\"created_at\":\"2026-02-09 07:13:43\",\"extension\":\"jpg\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":20314,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eLaterality concordance of AI (right vs left breast) among malignant examinations in the biopsy-referred cohort.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"4.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8701935/v1/fed6c87f14654bb79b00355b.jpg\"},{\"id\":104876806,\"identity\":\"1806a39a-71f8-4b9f-8d4d-a9230fec733e\",\"added_by\":\"auto\",\"created_at\":\"2026-03-18 08:43:43\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":1470456,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-8701935/v1/86835cfb-8250-41ba-8b2a-859ec1be9c06.pdf\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"Clinical performance tradeoffs of ChatGPT-5.2 Thinking (OpenAI) compared with radiologist interpretation in biopsy-referred mammography: cancer detection, false positives, and laterality\",\"fulltext\":[{\"header\":\"1. Introduction\",\"content\":\"\\u003cp\\u003eBreast cancer remains a leading cause of cancer-related morbidity and mortality among women worldwide and continues to place a substantial clinical and economic burden on health systems [1, 2]. Screening mammography has been central to early detection programs and has contributed to stage migration and improvements in survival through earlier diagnosis and timely referral to definitive therapy [3, 4]. Despite these gains, the performance of screening remains variable across programs and settings. Reader sensitivity and specificity depend on multiple factors that include breast density, lesion conspicuity, workload, double-reading policies, training, and access to prior images for comparison [5\\u0026ndash;8]. False negatives may delay diagnosis while false positives may trigger unnecessary callbacks, added imaging, anxiety, and biopsies that do not yield cancer [9\\u0026ndash;11]. These realities have motivated sustained efforts to enhance accuracy while preserving efficiency in high-volume screening environments [12, 13].\\u003c/p\\u003e \\u003cp\\u003eArtificial intelligence-based systems for mammography have emerged as a promising approach to support detection and decision making [14, 15]. Deep-learning methods trained on large datasets can learn visual patterns associated with masses, microcalcifications, architectural distortion, and asymmetries, and then produce image-level or case-level predictions that can be integrated into the reading workflow [16, 17]. Several studies report that AI can match or approach single-reader performance and can assist readers through triage, prioritization, or second reader-style support [18\\u0026ndash;21]. At the same time, the literature also documents failure modes that include difficulties with subtle calcifications, architectural distortion, and laterality or localization errors, as well as performance variability across vendors, acquisition protocols, and clinical settings [22\\u0026ndash;25]. There is an ongoing need for rigorous evaluation against robust reference standards and for transparent reporting of metrics that matter to patients and clinicians [26, 27]. In parallel, patients increasingly view their reports directly via portals, and a recent systematic review shows sustained patient demand for access alongside comprehension- and anxiety-related challenges; therefore, readability and patient-centred communication should be explicit design criteria when integrating AI outputs into screening workflows [28]. In this context, patient-facing AI interpretations could function as a structured \\u0026lsquo;second look\\u0026rsquo; explanation that helps patients formulate questions and participate in shared decision-making during clinical consultations. However, such use should be positioned as decision support that complements rather than replaces radiologist assessment and clinician-guided management\\u003c/p\\u003e \\u003cp\\u003eBiopsy confirmation offers a decisive reference to evaluate diagnostic systems because it anchors analyses to ground truth rather than to follow-up imaging alone [29]. Cohorts built on biopsy-confirmed cancers allow unbiased estimation of sensitivity and enable stratified analyses by lesion type, size, and breast density [29, 30]. When paired with an appropriate sample of biopsy-negative controls or verified benign outcomes, these cohorts also enable meaningful estimation of specificity and accuracy [30]. In addition, expert review of discordant cases can reveal systematic error patterns that inform future model development, calibration, and deployment strategies [31].\\u003c/p\\u003e \\u003cp\\u003eThe present study evaluates ChatGPT-5.2 Thinking (OpenAI) as a stand-alone model for the clinically defined imaging task of examination-level malignancy classification using bilateral CC and MLO screening mammography views in a biopsy-referred cohort with pathology reference. We hypothesized that the model would achieve sensitivity comparable to radiologists for biopsy-proven malignancy but would demonstrate lower specificity due to benign overcalling in a biopsy-enriched test set. We additionally assessed laterality performance and feature-type error patterns (mass, architectural distortion, microcalcifications, and parenchymal density) to identify clinically actionable failure modes [32].\\u003c/p\\u003e\"},{\"header\":\"2. Methods\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.1. Design and setting\\u003c/h2\\u003e \\u003cp\\u003e Retrospective, multicenter diagnostic-accuracy study conducted across several cities in Saudi Arabia within participating breast imaging centers. An upstream screening population of 1,225 women was accrued from multiple health systems during the study period. From this screened cohort, we constructed a biopsy-referred test set by including examinations that proceeded to image-guided or surgical biopsy based on clinical assessment. The analytic test set comprised 100 biopsy-referred mammography examinations (four standard 2D views per exam; 400 images total), including 61 biopsy-confirmed malignancies and 39 biopsy-negative controls, linked deterministically to each site\\u0026rsquo;s pathology database to establish a biopsy-anchored reference standard.\\u003c/p\\u003e \\u003cp\\u003eThis design was selected to ensure definitive ground truth for malignancy status and to enable direct estimation of sensitivity and specificity against pathology. However, because the cohort is enriched for biopsy-referred cases, it is not representative of a general screening population and may yield inflated sensitivity and depressed specificity relative to true screening prevalence; we therefore interpret performance as applicable to a biopsy-referred triage / concurrent-aid scenario rather than population screening. Accordingly, we quantified case-level sensitivity, specificity, and accuracy for malignant vs benign biopsy outcome (primary endpoint), and evaluated laterality performance and feature-type patterns (secondary endpoints).\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec4\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.2. Imaging and readers\\u003c/h2\\u003e \\u003cp\\u003eFor each case, bilateral craniocaudal (CC) and mediolateral oblique (MLO) views were exported as DICOM files and de-identified; each case comprised four images. Board-certified breast radiologists were blinded to pathology and to ChatGPT-5.2 Thinking outputs. Radiologists provided study-level BI-RADS\\u0026reg; assessment categories (0\\u0026ndash;5) and documented the suspected laterality (right vs left) of the abnormality prompting the assessment. Radiologist BI-RADS categories were used to generate binary calls at pre-specified thresholds for comparison with pathology and to classify true-positive, false-positive, true-negative, and false-negative outcomes. Laterality was used to define side-specific correctness among positive cases.\\u003c/p\\u003e \\u003cp\\u003eChatGPT-5.2 Thinking was evaluated in stand-alone mode using the same de-identified mammography views provided case-by-case (four images per examination) with a fixed, pre-specified BI-RADS\\u0026ndash;guided prompt (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e). The model generated an ordinal BI-RADS category (0\\u0026ndash;5) and suspected laterality for each case. ChatGPT-5.2 Thinking was not used to create or alter images, define ground truth, or compute statistical metrics; all performance calculations were performed by the study team using pathology as the reference standard.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec5\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.3. Artificial intelligence and prompting\\u003c/h2\\u003e \\u003cp\\u003eChatGPT-5.2 Thinking (OpenAI) was the only AI system evaluated in this study. The model was used in stand-alone mode to interpret de-identified bilateral CC and MLO mammography views. A fixed, pre-specified prompt instructed the model to produce: (i) an ordinal BI-RADS\\u0026reg; assessment category (0\\u0026ndash;5) using BI-RADS-consistent terminology, (ii) a binary malignancy classification derived from a pre-defined BI-RADS threshold, and (iii) suspected laterality (right vs left) when an abnormality was identified. The prompt was locked prior to evaluation and applied uniformly across cases (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e). ChatGPT-5.2 Thinking was not used to generate ground-truth labels, alter images, select cases, or compute performance statistics; all metrics were calculated by the study team using pathology as the reference standard.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec6\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.4. Statistical analysis, software, and reproducibility\\u003c/h2\\u003e \\u003cp\\u003eCategorical variables were summarized as counts and percentages, and continuous variables as means with standard deviations or medians with interquartile ranges, according to distributional shape. Primary case-level performance metrics were sensitivity, specificity, and overall accuracy; laterality accuracy (right vs left breast) was analyzed as a secondary endpoint. Exact 95% confidence intervals were calculated for proportions using the Clopper Pearson method, and paired proportions were compared with McNemar tests. Threshold-based summaries (e.g., BI-RADS\\u0026thinsp;\\u0026ge;\\u0026thinsp;3/\\u0026ge;4/\\u0026ge;5) are reported with corresponding true-positive and false-positive counts. Missing data were handled with complete-case analysis. All hypothesis tests were two-sided with α\\u0026thinsp;=\\u0026thinsp;0.05. Pre-specified subgroup analyses included lesion \\u0026ldquo;sign\\u0026rdquo; (mass, microcalcifications, architectural distortion, parenchymal density) and breast density categories when sample size permitted stable estimates.\\u003c/p\\u003e \\u003cp\\u003eAll analyses were executed in Python and R on secured institutional workstations. Reproducible pipelines were implemented as version-controlled scripts with fixed random seeds, where applicable; summary tables and figures were generated directly from analysis outputs. Parameter settings, analysis logs, and software versions were archived alongside the study protocol to facilitate replication. Institutional oversight was provided by the University of Wisconsin\\u0026ndash;Milwaukee IRB (IRB #20.230), with a waiver of informed consent for the use of de-identified, minimal-risk data. Data availability: De-identified imaging data cannot be publicly shared due to patient-privacy restrictions and institutional policy. Analysis code resides on an access-controlled, encrypted institutional workstation under the corresponding author\\u0026rsquo;s account and can be provided to qualified researchers upon reasonable request, subject to IRB oversight and a data-use agreement.\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"3. Results\",\"content\":\"\\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e3.1. Diagnostic Performance Comparison between Radiologists and AI\\u003c/h2\\u003e \\u003cp\\u003eFigure \\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e presents a comparative analysis of radiologist and AI diagnostic performance metrics, using biopsy results as the gold standard. The AI model demonstrated a higher sensitivity (95.08%) than radiologists (81.97%), indicating a stronger ability to identify biopsy-confirmed malignant cases. However, this improvement came at the expense of specificity, which dropped sharply to 10.26% compared to 56.41% for radiologists. Consequently, radiologists achieved a higher overall accuracy (72.00%) compared to AI (62.00%).\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eIn absolute terms, ChatGPT-5.2 correctly identified 58 true-positive cases but also generated 35 false-positive predictions, while radiologists recorded 50 true positives and 17 false positives. The AI missed only three true malignancies (FN\\u0026thinsp;=\\u0026thinsp;3), whereas radiologists missed eleven (FN\\u0026thinsp;=\\u0026thinsp;11). Although AI demonstrated better sensitivity, expert image review revealed that many of its false-positive detections corresponded to benign structures or peripheral artifacts, indicating misclassification rather than true lesion recognition.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec9\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e3.2. AI Performance by Radiologic Feature Type\\u003c/h2\\u003e \\u003cp\\u003eFigure \\u003cspan refid=\\\"Fig6\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003e summarizes the AI model\\u0026rsquo;s diagnostic metrics across four radiologic lesion categories: Mass, Architectural Distortion (AS), Microcalcifications (MC), and Parenchymal Density (PD). Among all lesion types, parenchymal density (PD) achieved the highest sensitivity (0.818), suggesting that AI performed best when detecting diffuse density changes. Mass lesions also demonstrated relatively high sensitivity (0.714) but suffered from extremely low specificity (0.125), reflecting a tendency of AI to misclassify normal dense areas as suspicious masses. Architectural distortion (AS) showed a balanced performance, with sensitivity of 0.455 and high specificity of 0.84 the most favorable trade-off among all categories. In contrast, microcalcifications (MC) were the weakest feature type for AI, with sensitivity of only 0.25 and precision of 0.211, indicating poor recognition of small-scale calcified foci. Overall, the pattern suggests that AI performs relatively well in identifying large-area abnormalities (PD, mass) but lacks reliability in localizing fine-grained structures (MC, AS). This variability across lesion types aligns with radiologists\\u0026rsquo; qualitative observations that AI often highlights regions adjacent to, rather than directly over, true lesions.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec10\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e3.3. AI Performance in Breast-Side Localization\\u003c/h2\\u003e \\u003cp\\u003eFigure \\u003cspan refid=\\\"Fig8\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e displays the AI\\u0026rsquo;s accuracy in correctly identifying the laterality (right vs. left breast) of detected abnormalities. The model correctly localized the disease in 37 cases (\\u0026ldquo;True side\\u0026rdquo;) and failed in 24 cases (\\u0026ldquo;False side\\u0026rdquo;). This yields an overall laterality accuracy of 60.66%, reflecting moderate spatial awareness but insufficient precision for clinical reliability. In many incorrect cases, AI successfully detected the presence of a lesion but marked it on the contralateral breast or outside the actual parenchymal area. Such errors underscore the model\\u0026rsquo;s limitations in spatial alignment and contextual awareness, particularly in cases with overlapping tissue density or bilateral findings.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"4. Discussion\",\"content\":\"\\u003cp\\u003eIn this biopsy-referred, pathology-anchored test set, ChatGPT-5.2 Thinking demonstrated higher sensitivity but markedly lower specificity than radiologists, resulting in higher false-positive prompting. Clinically, this pattern is consistent with a tool that may help reduce missed cancers if used as a concurrent aid (e.g., prompting a \\u0026ldquo;second look\\u0026rdquo; on cases otherwise read as benign), but it may also increase additional workup and patient anxiety due to benign overcalling. Beyond workflow impact, accessible AI outputs may also support patient engagement by helping patients understand the reason for further assessment and arrive to visits prepared for a more focused discussion. Any patient-facing use should remain tightly coupled to clinician review to avoid misinterpretation and unintended delays in care. This tradeoff is consistent with prior evaluations where deep learning systems increase case finding but introduce benign overcalls under real world mix of screening and diagnostic studies [1, 2, 4, 6, 11, 17, 22, 24]. Therefore, the most appropriate near-term clinical framing is triage/prioritization or concurrent-aid support, not replacement of radiologist interpretation [5, 10, 15, 20].\\u003c/p\\u003e \\u003cp\\u003eFeature level behavior clarifies where artificial intelligence helped and where it struggled. Parenchymal density produced the highest sensitivity, which aligns with studies showing that textural and global parenchymal patterns are well captured by convolutional and transformer-based encoders trained at image scale [7, 13, 18]. Architectural distortion achieved the best specificity but only modest sensitivity, a conservative posture that mirrors prior reports in which models avoid distortion overcalls yet miss subtle tethering or spiculations in dense tissue [9, 14, 18]. Mass showed moderate sensitivity but very low specificity, consistent with the known risk of mistaking overlapping tissue and benign masses for malignancy, especially in heterogeneously dense breasts [3, 5, 19, 29]. Microcalcifications remained the weakest category, echoing literature that small, sparse, and high frequency signals are sensitive to pixel resolution, preprocessing, and training set curation [8, 12, 23, 27]. Together these patterns suggest practical parameterization: use conservative thresholds when calcifications or mass like appearances are suspected, while allowing artificial intelligence more freedom to surface diffuse parenchymal abnormalities [10, 13, 18].\\u003c/p\\u003e \\u003cp\\u003eLaterality accuracy was 60.7%, which represents a meaningful constraint for workflow. Wrong side assignment can lead to additional views, unnecessary callbacks, and avoidable anxiety, even when the case is correctly classified as positive [3, 16, 26]. Multiple groups have improved side awareness by enforcing cross view consistency checks and by incorporating quadrant aware attention maps or side conditioned decoders [7, 17, 31]. Our results support adoption of such side aware constraints before routine deployment. Breast density influenced both artificial intelligence and reader performance. Sensitivity decreased in heterogeneously and extremely dense categories, consistent with lesion masking from x ray attenuation and reduced conspicuity [1, 4, 27]. Studies that integrate prior examinations, tomosynthesis, or targeted ultrasound have documented partial recovery of sensitivity, suggesting that artificial intelligence systems which fuse priors or cross modality context may mitigate density related loss [6, 17, 29]. Until such fusion is widely available, cautious interpretation in dense breasts and targeted secondary imaging remain necessary.\\u003c/p\\u003e \\u003cp\\u003eComparison with published benchmarks indicates that our radiologist accuracy aligns with contemporary single reader performance, while the artificial intelligence profile falls within the range reported for research grade systems evaluated outside the original development distribution [2, 11, 17, 24]. Importantly, several multi center studies have shown that artificial intelligence specificity is most unstable when prevalence, vendor mix, and positioning differ from training data [9, 14, 30, 33, 34]. This reinforces the value of local calibration prior to adoption.\\u003c/p\\u003e \\u003cp\\u003eAs an operational model, a concurrent-aid approach elevates higher-risk studies in the worklist and cues focused assessment yet maintains independent judgment [5, 15, 20, 35, 36]. To reduce benign overcalls, some programs apply decision curves that trade small losses in sensitivity for meaningful reductions in unnecessary recalls [2, 24, 37]. Artificial intelligence as an independent second reader may substitute for double reading in resource constrained settings only after specificity and laterality accuracy are improved and clear escalation rules for discordant cases are validated prospectively [4, 22, 30]. User interface design remains critical to avoid automation bias and to protect the benefits of independent reads [10, 19].\\u003c/p\\u003e \\u003cp\\u003eSafety and governance require noting that, while laboratory anchoring reduces verification bias for malignancy, benign ground truth still partly depends on imaging stability, which can skew specificity estimates [1, 35, 38]. Prospective quality assurance should include periodic review of false positives and false negatives with feedback loops to both model and readers, an approach associated with measurable gains in other imaging domains [18, 39, 40]. Transparent reporting with confidence intervals, calibration plots, and decision curves enables oversight bodies to judge tradeoffs that reflect local priorities [2, 29, 41].\\u003c/p\\u003e \\u003cp\\u003eEquity and generalizability require attention, as documented performance drift across vendors, compression paddles, and population subgroups can affect both sensitivity and specificity [9, 14, 30]. External validation across diverse sites with predefined acceptability margins and periodic recalibration after deployment is therefore essential [17, 30, 31]. Adherence to reporting standards and public documentation of thresholds and update cadence support reproducibility and fair comparison across systems [19, 29]. Implications for clinical practice: Given its higher sensitivity but lower specificity, artificial intelligence is best used to trigger second looks on studies otherwise read as negative, helping reduce missed cancers [5, 10]. However, because benign overcalls carry cost and anxiety, programs should track recall rate, positive predictive value of biopsy recommendation, and time to diagnostic resolution after adoption [2, 22, 24]. The feature type profile suggests that readers should be most skeptical of artificial intelligence prompts for mass in dense tissue and for subtle calcifications, while considering artificial intelligence prompts for diffuse parenchymal patterns as higher yield [8, 13, 18].\\u003c/p\\u003e \\u003cp\\u003eTo boost specificity, augment training sets with common benign look-alikes including tissue overlap, fibroadenomas with calcifications, and postsurgical change while employing cost-sensitive losses that penalize false positives [8, 12, 19]. Second, improve calcification detection by increasing native pixel resolution, adding calcification specific augmentations, and using multi scale heads tuned for small object detection [8, 23, 27]. Third, harden laterality by enforcing side aware constraints and cross view agreement, and by rejecting outputs that conflict with view geometry [3, 16, 26, 31]. Fourth, explore fusion of priors and tomosynthesis to mitigate density effects and improve robustness across acquisition variability [6, 17, 29, 30].\\u003c/p\\u003e\"},{\"header\":\"5. Conclusion\",\"content\":\"\\u003cp\\u003eArtificial intelligence approached radiologist level detection but had lower specificity and therefore lower overall accuracy. Performance differed by feature type, with strengths in PD, conservative high specificity for AD, overcalling in MAS, and the weakest performance in MC. Laterality accuracy was modest at sixty-point seven percent, highlighting a practical barrier to deployment. These findings support artificial intelligence as a concurrent aid or prioritization tool rather than a stand-alone reader, and motivate targeted refinement in specificity, microcalcification handling, and breast side assignment, followed by prospective and external validation.\\u003c/p\\u003e\"},{\"header\":\"Abbreviations\",\"content\":\"\\u003cp\\u003eAI, artificial intelligence\\u003c/p\\u003e\\n\\u003cp\\u003eBI-RADS, Breast Imaging Reporting and Data System\\u003c/p\\u003e\\n\\u003cp\\u003eCC, craniocaudal\\u003c/p\\u003e\\n\\u003cp\\u003eMLO, mediolateral oblique\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\\u003c/p\\u003e\\u003cp\\u003e \\u003ch2\\u003eCompeting Interests\\u003c/h2\\u003e \\u003cp\\u003eThe authors declare no relevant financial or non-financial interests related to this work.\\u003c/p\\u003e \\u003c/p\\u003e\\u003cp\\u003e \\u003ch2\\u003eEthics approval and consent to participate\\u003c/h2\\u003e \\u003cp\\u003e Ethics approval and consent to participate: The study was approved by the University of Wisconsin\\u0026ndash;Milwaukee Institutional Review Board (IRB #20.230), and was conducted in accordance with the Declaration of Helsinki and relevant institutional guidelines and regulations; informed consent was waived by the IRB because the study used de-identified, minimal-risk retrospective data.\\u003c/p\\u003e \\u003c/p\\u003e\\u003ch2\\u003eFunding\\u003c/h2\\u003e \\u003cp\\u003eThis work was supported by King Saud University. Mohammad Alarifi has received research support from King Saud University (KSU).\\u003c/p\\u003e\\u003ch2\\u003eAuthor Contribution\\u003c/h2\\u003e\\u003cp\\u003eConceptualization: Alarifi, AlmanaaMethodology: Alarifi, Jabour, AbanomyValidation: Aloufi, AlahmadFormal analysis: Alarifi, Jabour, AbanomyInvestigation: Aloufi, Abanomy, Alahmad, Alenazi, AlmanaaResources: Alenazi, AlmanaaData curation: Aloufi, JabourVisualization: Alarifi, AlshediWriting \\u0026ndash; original draft: Alarifi, Aloufi, JabourWriting \\u0026ndash; review \\u0026amp; editing: All authorsSupervision: Alarifi, AlmanaaProject administration: AlarifiAll authors read and approved the final manuscript.\\u003c/p\\u003e\\u003ch2\\u003eData Availability\\u003c/h2\\u003e\\u003cp\\u003eDe-identified imaging data cannot be publicly shared due to patient-privacy restrictions and institutional policy. Analysis code resides on an access-controlled, encrypted institutional workstation under the corresponding author\\u0026rsquo;s account and can be provided to qualified researchers upon reasonable request, subject to IRB oversight and a data-use agreement.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\n\\u003cli\\u003eRodr\\u0026iacute;guez-Ruiz A, Krupinski E, Mordang J-J, Schilling K, Heywang-K\\u0026ouml;brunner SH, Sechopoulos I, et al \\u003cstrong\\u003e(2019)\\u003c/strong\\u003e Detection of breast cancer with mammography: effect of an artificial intelligence support system. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e290(2):305\\u0026ndash;314\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eResch D, Lo Gullo R, Teuwen J, Semturs F, Hummel J, Resch A, et al \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e AI-enhanced mammography with digital breast tomosynthesis for breast cancer detection: clinical value and comparison with human performance. \\u003cstrong\\u003eRadiol Imaging Cancer\\u003c/strong\\u003e\\u003cstrong\\u003e6(4):e230149\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003evan Winkel SL, Rodr\\u0026iacute;guez-Ruiz A, Appelman L, Gubern-M\\u0026eacute;rida A, Karssemeijer N, Teuwen J, et al \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Impact of artificial intelligence support on accuracy and reading time in breast tomosynthesis image interpretation: a multi-reader multi-case study. \\u003cstrong\\u003eEur Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e31(11):8682\\u0026ndash;8691\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003ePark EK, Kwak S, Lee W, Choi JS, Kooi T, Kim E-K \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Impact of AI for digital breast tomosynthesis on breast cancer detection and interpretation time. \\u003cstrong\\u003eRadiol Artif Intell\\u003c/strong\\u003e\\u003cstrong\\u003e6(3):e230318\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eHadadi I, Rae W, Clarke J, McEntee M, Ekpo E \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Breast cancer detection: comparison of digital mammography and digital breast tomosynthesis across non-dense and dense breasts. \\u003cstrong\\u003eRadiography\\u003c/strong\\u003e\\u003cstrong\\u003e27(4):1027\\u0026ndash;1032\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLauritzen AD, Rodr\\u0026iacute;guez-Ruiz A, von Euler-Chelpin MC, Lynge E, Vejborg I, Nielsen M, et al \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e An artificial intelligence\\u0026ndash;based mammography screening protocol for breast cancer: outcome and radiologist workload. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e304(1):41\\u0026ndash;49\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eKim H-E, Kim HH, Han B-K, Kim KH, Han K, Nam H, et al \\u003cstrong\\u003e(2020)\\u003c/strong\\u003e Changes in cancer detection and false-positive recall in mammography using artificial intelligence: a retrospective, multireader study. \\u003cstrong\\u003eLancet Digit Health\\u003c/strong\\u003e\\u003cstrong\\u003e2(3):e138\\u0026ndash;e148\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eRodriguez-Ruiz A, L\\u0026aring;ng K, Gubern-Merida A, Broeders M, Gennaro G, Clauser P, et al \\u003cstrong\\u003e(2019)\\u003c/strong\\u003e Stand-alone artificial intelligence for breast cancer detection in mammography: comparison with 101 radiologists. \\u003cstrong\\u003eJ Natl Cancer Inst\\u003c/strong\\u003e\\u003cstrong\\u003e111(9):916\\u0026ndash;922\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eChoi WJ, An JK, Woo JJ, Kwak HY \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Comparison of diagnostic performance in mammography assessment: radiologist with reference to clinical information versus standalone artificial intelligence detection. \\u003cstrong\\u003eDiagnostics (Basel)\\u003c/strong\\u003e\\u003cstrong\\u003e13(1):117\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLee SE, Hong H, Kim E-K \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Diagnostic performance with and without artificial intelligence assistance in real-world screening mammography. \\u003cstrong\\u003eEur J Radiol Open\\u003c/strong\\u003e\\u003cstrong\\u003e12:100545\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003ePacil\\u0026egrave; S, Lopez J, Chone P, Bertinotti T, Grouin JM, Fillard P \\u003cstrong\\u003e(2020)\\u003c/strong\\u003e Improving breast cancer detection accuracy of mammography with the concurrent use of an artificial intelligence tool. \\u003cstrong\\u003eRadiol Artif Intell\\u003c/strong\\u003e\\u003cstrong\\u003e2(6):e190208\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003ede Vries CF, Colosimo SJ, Staff RT, Dymiter JA, Yearsley J, Dinneen D, et al \\u003cstrong\\u003e(2023)\\u003c/strong\\u003e Impact of different mammography systems on artificial intelligence performance in breast cancer screening. \\u003cstrong\\u003eRadiol Artif Intell\\u003c/strong\\u003e\\u003cstrong\\u003e5(3):e220146\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eKim YS, Jang M-J, Lee SH, Kim S-Y, Ha SM, Kwon BR, et al \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Use of artificial intelligence for reducing unnecessary recalls at screening mammography: a simulation study. \\u003cstrong\\u003eKorean J Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e23(12):1241\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAnderson AW, Marinovich ML, Houssami N, Lowry KP, Elmore JG, Buist DS, et al \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Independent external validation of artificial intelligence algorithms for automated interpretation of screening mammography: a systematic review. \\u003cstrong\\u003eJ Am Coll Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e19(2):259\\u0026ndash;273\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eMansour S, Sweed E, Gomaa MMM, Hussein SA, Abdallah E, Nada YM, et al \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Artificial intelligence as an initial reader for double reading in breast cancer screening: a prospective initial study of 32,822 mammograms of the Egyptian population. \\u003cstrong\\u003eEgypt J Radiol Nucl Med\\u003c/strong\\u003e\\u003cstrong\\u003e55(1):181\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLee JH, Kim KH, Lee EH, Ahn JS, Ryu JK, Park YM, et al \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Improving the performance of radiologists using artificial intelligence-based detection support software for mammography: a multi-reader study. \\u003cstrong\\u003eKorean J Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e23(5):505\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eNanaa M, Gupta VO, Hickman SE, Allajbeu I, Payne NR, Arponen O, et al \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Accuracy of an artificial intelligence system for interval breast cancer detection at screening mammography. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e312(2):e232303\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLarsen M, Olstad CF, Lee CI, Hovda T, Hoff SR, Martiniussen MA, et al \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Performance of an artificial intelligence system for breast cancer detection on screening mammograms from BreastScreen Norway. \\u003cstrong\\u003eRadiol Artif Intell\\u003c/strong\\u003e\\u003cstrong\\u003e6(3):e230375\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eL\\u0026aring;ng K, Hofvind S, Rodr\\u0026iacute;guez-Ruiz A, Andersson I \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Can artificial intelligence reduce the interval cancer rate in mammography screening? \\u003cstrong\\u003eEur Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e31(8):5940\\u0026ndash;5947\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLarsen M, Aglen CF, Hoff SR, Lund-Hanssen H, Hofvind S \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Possible strategies for use of artificial intelligence in screen-reading of mammograms, based on retrospective data from 122,969 screening examinations. \\u003cstrong\\u003eEur Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e32(12):8238\\u0026ndash;8246\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAlarifi M \\u003cstrong\\u003e(2025)\\u003c/strong\\u003e Radiologists\\u0026rsquo; views on AI and the future of radiology: insights from a US national survey. \\u003cstrong\\u003eBr J Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e:tqaf222\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eZeng A, Houssami N, Noguchi N, Nickel B, Marinovich ML \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Frequency and characteristics of errors by artificial intelligence in reading screening mammography: a systematic review. \\u003cstrong\\u003eBreast Cancer Res Treat\\u003c/strong\\u003e\\u003cstrong\\u003e207(1):1\\u0026ndash;13\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAbuAbeileh EA, Abuabeileh Z, AbuAbeileh Z \\u003cstrong\\u003e(2025)\\u003c/strong\\u003e Screening mammography and artificial intelligence: a comprehensive systematic review. \\u003cstrong\\u003eCureus\\u003c/strong\\u003e\\u003cstrong\\u003e17(2):e\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eL\\u0026aring;ng K, Dustler M, Dahlblom V, \\u0026Aring;kesson A, Andersson I, Zackrisson S \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Identifying normal mammograms in a large screening population using artificial intelligence. \\u003cstrong\\u003eEur Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e31(3):1687\\u0026ndash;1692\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAlarifi M \\u003cstrong\\u003e(2025)\\u003c/strong\\u003e Appropriateness of thyroid nodule cancer risk assessment and management recommendations provided by large language models. \\u003cstrong\\u003eJ Imaging Inform Med\\u003c/strong\\u003e\\u003cstrong\\u003e:1\\u0026ndash;12\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eUematsu T, Nakashima K, Harada TL, Nasu H, Igarashi T \\u003cstrong\\u003e(2023)\\u003c/strong\\u003e Comparisons between artificial intelligence computer-aided detection synthesized mammograms and digital mammograms when used alone and in combination with tomosynthesis images in a virtual screening setting. \\u003cstrong\\u003eJpn J Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e41(1):63\\u0026ndash;70\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eLee SE, Yoon JH, Son N-H, Han K, Moon HJ \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Screening in patients with dense breasts: comparison of mammography, artificial intelligence, and supplementary ultrasound. \\u003cstrong\\u003eAJR Am J Roentgenol\\u003c/strong\\u003e\\u003cstrong\\u003e222(1):e2329655\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAlarifi M, Hughes MC, Jabour AM, Alashban Y, Vernon E \\u003cstrong\\u003e(2024)\\u003c/strong\\u003e Patient, referring physician, and radiologist opinions over time on providing patients access to radiology reports: a systematic review. \\u003cstrong\\u003eJ Am Coll Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e21(12):1862\\u0026ndash;1874\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eDembrower K, Crippa A, Col\\u0026oacute;n E, Eklund M, Strand F \\u003cstrong\\u003e(2023)\\u003c/strong\\u003e Artificial intelligence for breast cancer detection in screening mammography in Sweden: a prospective, population-based, paired-reader, non-inferiority study. \\u003cstrong\\u003eLancet Digit Health\\u003c/strong\\u003e\\u003cstrong\\u003e5(10):e703\\u0026ndash;e711\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eRomero-Mart\\u0026iacute;n S, El\\u0026iacute;as-Cabot E, Raya-Povedano JL, Gubern-M\\u0026eacute;rida A, Rodr\\u0026iacute;guez-Ruiz A, \\u0026Aacute;lvarez-Benito M \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Stand-alone use of artificial intelligence for digital mammography and digital breast tomosynthesis screening: a retrospective evaluation. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e302(3):535\\u0026ndash;542\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eWatanabe AT, Lim V, Vu HX, Chim R, Weise E, Liu J, et al \\u003cstrong\\u003e(2019)\\u003c/strong\\u003e Improved cancer detection using artificial intelligence: a retrospective evaluation of missed cancers on mammography. \\u003cstrong\\u003eJ Digit Imaging\\u003c/strong\\u003e\\u003cstrong\\u003e32(4):625\\u0026ndash;637\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAlarifi M \\u003cstrong\\u003e(2025)\\u003c/strong\\u003e Exploring radiologists\\u0026rsquo; use of AI chatbots for assistance in image interpretation: patterns of use and trust evaluation. \\u003cstrong\\u003eJ Imaging Inform Med\\u003c/strong\\u003e\\u003cstrong\\u003e:1\\u0026ndash;10\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eAlabousi M, Wadera A, Al-Ghita MK, Al-Ghetaa RK, Salameh J-P, Pozdnyakov A, et al \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Performance of digital breast tomosynthesis, synthetic mammography, and digital mammography in breast cancer screening: a systematic review and meta-analysis. \\u003cstrong\\u003eJ Natl Cancer Inst\\u003c/strong\\u003e\\u003cstrong\\u003e113(6):680\\u0026ndash;690\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eChikarmane SA, Cochon LR, Khorasani R, Sahu S, Giess CS \\u003cstrong\\u003e(2021)\\u003c/strong\\u003e Screening mammography performance metrics of 2D digital mammography versus digital breast tomosynthesis in women with a personal history of breast cancer. \\u003cstrong\\u003eAJR Am J Roentgenol\\u003c/strong\\u003e\\u003cstrong\\u003e217(3):587\\u0026ndash;594\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eDembrower K, W\\u0026aring;hlin E, Liu Y, Salim M, Smith K, Lindholm P, et al \\u003cstrong\\u003e(2020)\\u003c/strong\\u003e Effect of artificial intelligence-based triaging of breast cancer screening mammograms on cancer detection and radiologist workload: a retrospective simulation study. \\u003cstrong\\u003eLancet Digit Health\\u003c/strong\\u003e\\u003cstrong\\u003e2(9):e468\\u0026ndash;e474\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eRodriguez-Ruiz A, L\\u0026aring;ng K, Gubern-Merida A, Teuwen J, Broeders M, Gennaro G, et al \\u003cstrong\\u003e(2019)\\u003c/strong\\u003e Can we reduce the workload of mammographic screening by automatic identification of normal exams with artificial intelligence? a feasibility study. \\u003cstrong\\u003eEur Radiol\\u003c/strong\\u003e\\u003cstrong\\u003e29(9):4825\\u0026ndash;4832\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eDang L-A, Chazard E, Poncelet E, Serb T, Rusu A, Pauwels X, et al \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Impact of artificial intelligence in breast cancer screening with mammography. \\u003cstrong\\u003eBreast Cancer\\u003c/strong\\u003e\\u003cstrong\\u003e29(6):967\\u0026ndash;977\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eSasaki M, Tozaki M, Rodr\\u0026iacute;guez-Ruiz A, Yotsumoto D, Ichiki Y, Terawaki A, et al \\u003cstrong\\u003e(2020)\\u003c/strong\\u003e Artificial intelligence for breast cancer detection in mammography: experience of use of the ScreenPoint Medical Transpara system in 310 Japanese women. \\u003cstrong\\u003eBreast Cancer\\u003c/strong\\u003e\\u003cstrong\\u003e27(4):642\\u0026ndash;651\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eHadadi I, Rae W, Clarke J, McEntee M, Ekpo E \\u003cstrong\\u003e(2022)\\u003c/strong\\u003e Breast cancer detection across dense and non-dense breasts: markers of diagnostic confidence and efficacy. \\u003cstrong\\u003eActa Radiol Open\\u003c/strong\\u003e\\u003cstrong\\u003e11(1):20584601211072279\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eKolb TM, Lichy J, Newhouse JH \\u003cstrong\\u003e(2002)\\u003c/strong\\u003e Comparison of the performance of screening mammography, physical examination, and breast US and evaluation of factors that influence them: an analysis of 27,825 patient evaluations. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e225(1):165\\u0026ndash;175\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003cli\\u003eYoon JH, Strand F, Baltzer PA, Conant EF, Gilbert FJ, Lehman CD, et al \\u003cstrong\\u003e(2023)\\u003c/strong\\u003e Standalone AI for breast cancer detection at screening digital mammography and digital breast tomosynthesis: a systematic review and meta-analysis. \\u003cstrong\\u003eRadiology\\u003c/strong\\u003e\\u003cstrong\\u003e307(5):e222639\\u003c/strong\\u003e.\\u003c/li\\u003e\\n\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":true,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true},\"keywords\":\"Artificial intelligence, Mammography, Breast cancer, BI-RADS, Diagnostic accuracy, Reader study\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-8701935/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-8701935/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003ch2\\u003ePurpose\\u003c/h2\\u003e \\u003cp\\u003eTo compare ChatGPT-5.2 Thinking (OpenAI) with practicing radiologists for the clinically relevant, examination-level task of classifying biopsy-proven malignancy in a biopsy-referred mammography test set, and to assess performance by radiologic feature type and laterality\\u003c/p\\u003e\\u003ch2\\u003eMethods\\u003c/h2\\u003e \\u003cp\\u003eIn this multicenter retrospective study across several cities in Saudi Arabia, screening mammograms from an initial cohort of 1,225 women were linked to pathology to create a biopsy-anchored reference standard. Board-certified breast radiologists, blinded to pathology and model outputs, provided study-level BI-RADS\\u0026reg; assessments. ChatGPT-5.2 Thinking received de-identified bilateral CC and MLO views with a fixed BI-RADS\\u0026ndash;based prompt and produced an ordinal BI-RADS category (0\\u0026ndash;5) and suspected laterality. The analytic test set included 100 examinations that proceeded to biopsy (61 biopsy-confirmed cancers and 39 biopsy-negative controls). Primary outcomes were case-level sensitivity, specificity, and accuracy; secondary outcomes included laterality performance and feature-level patterns. All analyses were executed in Python and R on secured institutional workstations.\\u003c/p\\u003e\\u003ch2\\u003eResults\\u003c/h2\\u003e \\u003cp\\u003eChatGPT-5.2 demonstrated higher sensitivity than radiologists (95.1% vs 82.0%) but lower specificity (10.3% vs 56.4%), yielding lower overall accuracy (62.0% vs 72.0%). Feature-wise, AI showed highest sensitivity with dense parenchymal patterns and highest specificity for architectural distortion, tended to overcall mass-like findings, and performed weakest for microcalcifications. Laterality accuracy was 60.7%.\\u003c/p\\u003e\\u003ch2\\u003eConclusion\\u003c/h2\\u003e \\u003cp\\u003eIn this biopsy-referred, pathology-anchored evaluation, ChatGPT-5.2 Thinking showed higher sensitivity but substantially lower specificity than radiologists, supporting its potential role as a concurrent aid/triage signal rather than a stand-alone reader, pending prospective validation.\\u003c/p\\u003e\",\"manuscriptTitle\":\"Clinical performance tradeoffs of ChatGPT-5.2 Thinking (OpenAI) compared with radiologist interpretation in biopsy-referred mammography: cancer detection, false positives, and laterality\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2026-02-09 07:13:06\",\"doi\":\"10.21203/rs.3.rs-8701935/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"c9b48333-c75d-4a69-8889-0b148540cde0\",\"owner\":[],\"postedDate\":\"February 9th, 2026\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[],\"tags\":[],\"updatedAt\":\"2026-03-18T08:42:07+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2026-02-09 07:13:06\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-8701935\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-8701935\",\"identity\":\"rs-8701935\",\"version\":[\"v1\"]},\"buildId\":\"XKTyCvWXoU3ODBz1xrDgd\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}