Classifying diagnostic pitfalls in juxtapleural CT lesions: A root cause Authors: Kunquan Lai, Xinhuan Huang, Huan Lai, Huixian Chen, Suiqiao Huang, Honggang Xu, Yongshun Wu Published: 2026 ABSTRACT ObjectivesTo develop and assess the feasibility of a root-cause analysis (RCA) framework for classifying diagnostic pitfalls in the computed tomography (CT) interpretation of juxtapleural lesions.MethodsThis single-center, retrospective, exploratory taxonomy-development study was approved by our Institutional Review Board with a waiver of informed consent. To focus on the underlying mechanisms of diagnostic failure rather than overall institutional performance, we analyzed a cohort deliberately enriched for diagnostic difficulty. We analyzed 20 patients (mean age, 55.1 ± 16.9 years; 11 women) from a county-level hospital whose examinations were performed between December 2021 and March 2024. All patients had a definitive diagnosis established by histopathology/microbiology (n = 18) or conclusive clinical diagnosis (n = 2). Non-concordant cases (n = 18) were subjected to a structured RCA. Two radiologists independently categorized errors into three pitfall types.ResultsIn this difficulty-enriched, intentionally selected cohort—which is not representative of general diagnostic performance—major or moderate diagnostic discordance was identified in 18 of 20 cases (90.0%; 95% CI: 68.3%, 98.8%). These errors converged into three distinct pitfalls: (1) Interpretive Errors due to Misleading Features, representing the most common type (12/18, 66.7%); (2) Perceptual Errors (3/18, 16.7%); and (3) Errors of Integration (3/18, 16.7%). Inter-rater reliability for this categorization showed substantial agreement (Cohen's κ = 0.68; 95% CI: 0.41, 0.95), though the wide confidence interval reflects the limited sample size.ConclusionIn diagnostically challenging scenarios, errors in the CT interpretation of juxtapleural lesions exhibit systematic and classifiable patterns. Our novel three-pitfall taxonomy provides a preliminary yet promising framework for analyzing these errors, demonstrating feasibility and initial inter-rater reliability. External validation in larger, multicenter, prospective studies is required before this framework can be recommended for clinical implementation, educational deployment, or assessment of impact on diagnostic performance. Introduction Pleural and juxtapleural diseases encompass a wide spectrum of pathologies, presenting a significant diagnostic challenge in clinical practice [1], [2]. While chest radiography is often the initial imaging modality, computed tomography (CT) has become the cornerstone for further characterization [3]. However, a fundamental limitation persists: the considerable overlap in CT findings among disparate conditions. Studies like that by Piacibello et al. have highlighted the ongoing challenge of correlating CT features with definitive thoracoscopic diagnoses, especially in diffuse pleural diseases [4]. Malignant pleural mesothelioma (MPM), metastatic pleural disease (MPD), infectious processes like empyema, and even rare entities such as neurogenic tumors or sarcomas can exhibit remarkably similar imaging features [5], [6], [7], [8], [9]. This diagnostic ambiguity complicates clinical management and poses a critical risk to patient safety. The challenge of diagnostic accuracy in radiology is amplified by inherent human factors. Decades of research in patient safety and cognitive science have established that diagnostic errors are often systematic and predictable [10], [11]. A substantial body of evidence indicates that a majority of these errors are perceptual or cognitive in nature [12], [13]. Cognitive biases, such as "satisfaction of search," are particularly pernicious in interpreting thoracic images [11], [14], with mechanistic underpinnings rooted in a dynamic "expectancy bias" [15], [16]. These pressures are further exacerbated by systemic factors like work overload, which has been quantitatively shown to increase diagnostic error rates [17]. Existing literature has predominantly cataloged imaging features [6], [7], offering limited insight into the mechanisms of diagnostic failure. While prior studies, including the foundational work of Graber et al. [12] and Lee et al. [20], have explored radiological errors broadly and proposed general classification schemes, these frameworks were designed for generic diagnostic error categorization across all medical specialties and are not specifically tailored to the unique cognitive and perceptual challenges of thoracic CT interpretation of juxtapleural lesions. To our knowledge, no study has applied a systematic RCA framework grounded in cognitive science to specifically classify diagnostic pitfalls for juxtapleural lesions—a methodological gap that this exploratory study aims to address. We hypothesized that diagnostic errors in the interpretation of juxtapleural lesions are not random but are systematic and can be classified into a predictable taxonomy of pitfalls using an RCA framework. The aims of this exploratory study were therefore twofold: (1) to develop and categorize diagnostic errors in juxtapleural CT interpretation into a reproducible taxonomy of pitfalls using an adapted RCA framework; and (2) to perform a preliminary assessment of the inter-rater reliability of this classification framework as an initial step toward broader validation. Materials and methods Study design and ethical compliance This single-center, retrospective observational study was approved by the Institutional Review Board(IRB) of our Hospital (Anonymized Reference Number). Ethical approval was obtained prior to data analysis for this retrospective review of patient records. The board granted a waiver of written informed consent due to the retrospective design and use of fully anonymized data. The study was performed in compliance with the STROBE guidelines, and we adhered to the STARD 2015 guidelines where applicable [18]. Patient population A search of our institutional database for the period of December 2020 to July 2025 identified a consecutive cohort of 91 potentially eligible patients with juxtapleural lesions on chest CT. Inclusion criteria were: (a) adult patient (age ≥ 18 years) with a thoracic CT demonstrating a newly diagnosed juxtapleural lesion; and (b) availability of a definitive reference standard. After applying all inclusion and exclusion criteria (Fig. 1), a final study cohort of 20 patients, whose examinations were performed between December 2021 and March 2024, was established. Exclusion criteria are detailed in the patient selection flowchart (Fig. 1). Critically, the exclusion of cases with pathognomonic features or no significant diagnostic dilemma (n = 31) was a deliberate and pre-specified methodological decision, not a study limitation. This enrichment strategy was adopted because the study objective was not to measure institutional diagnostic accuracy, but to identify and characterize the cognitive mechanisms underlying diagnostic failure. By focusing on diagnostically challenging cases, we maximized the yield of informative errors for root cause analysis, analogous to the established methodological approach of case-enrichment used in cognitive error research [12], [19].Fig. 1Flowchart of patient selection for the study, according to the STARD 2015 guidelines. Flowchart of patient selection for the study, according to the STARD 2015 guidelines. The reference standard was established by histopathological or microbiological confirmation (n = 18) or a conclusive clinical diagnosis (n = 2). Conclusive clinical diagnosis required unanimous agreement by a multidisciplinary team of at least three experts, based on a combination of characteristic imaging evolution over a minimum of 12 months follow-up and irrefutable laboratory or clinical findings. This standard was applied in cases where invasive procedures were contraindicated or deemed not feasible. CT acquisition, reconstruction, and analysis All examinations were performed on one of two multidetector CT systems (Optima CT680 or Revolution CT; GE Healthcare). The institutional protocol included unenhanced imaging followed by tri-phasic contrast-enhanced scanning. Arterial phase imaging was initiated using a bolus-tracking technique, with subsequent venous and delayed phases acquired at standardized intervals (e.g., 40 s and 80 s for thoracic scans). A non-ionic iodinated contrast medium was administered intravenously (3–4 mL/s). Acquisition parameters were standardized, utilizing automatic tube current modulation to adhere to the ALARA principle. Images were reconstructed with a slice thickness of 1.25 mm using a standard soft-tissue kernel. Image analysis was conducted on GE Advantage Workstations (versions 4.6 and 4.7). All images were reviewed on institutional PACS (Jiangxi Zhongtong Bit or Mindray) with standardized window settings for mediastinum (width/level, 350/40 HU) and lung (1600/–550 HU). Radiation dose was managed according to the ALARA (As Low As Reasonably Achievable) principle, with both scanners utilizing automatic tube current modulation. While specific dose metrics were not systematically analyzed for this retrospective study, a qualitative review by senior radiologists confirmed that there were no systematic differences in image noise or lesion conspicuity between the two systems that would be expected to impact diagnostic interpretation. Image interpretation and root cause analysis The initial clinical CT reports, issued by staff radiologists at Shicheng County People's Hospital, served as the index test. For the core analysis, two board-certified radiologists (X.H., 10 years of experience; H.C., 6 years of experience) served as independent reviewers. Prior to the study, both reviewers underwent a calibration session using five test cases (not included in the final cohort) to ensure consistent application of the concordance criteria and the RCA scoring rubric. This session also confirmed their interpretation consistency and diagnostic workflow efficiency across both PACS systems used during the study period. During the formal review, the reviewers were blinded to the original reports and patient identifiers. Because root cause analysis requires comparison against a definitive diagnosis to trace the causal chain of error, the reviewers had access to the reference standard (final diagnosis) during the RCA phase; however, concordance categorization was performed independently of knowledge of the original report. The reviewers independently categorized the concordance between the index test and the reference standard. For all non-concordant cases, defined as those with either discordant (incorrect) or moderately concordant (partially correct) diagnoses (n = 18), a structured RCA was performed. Moderately concordant cases were included in the RCA because the failure to achieve a precise etiological diagnosis, even when the benign/malignant nature was correct, was hypothesized to stem from the same underlying systemic or cognitive pitfalls as major errors. We adapted the RCA framework proposed by Graber et al. [19]. To tailor this general framework specifically for radiological interpretation errors, we distilled it into five core domains relevant to the diagnostic process. The selection and refinement of these five domains followed a structured, multi-step process: first, we conducted a focused literature review of cognitive error taxonomies in radiology [10], [12], [13]; second, three senior thoracic radiologists (>20, 10, and 6 years of experience, respectively) independently proposed domain candidates; and third, consensus was reached through two rounds of iterative Delphi-style discussion, prioritizing domains that were (a) clinically distinct, (b) actionable, and (c) collectively covering the entire diagnostic pathway from image perception to report generation. This process resulted in a focused and reproducible scoring tool for our study. These domains were scored on a 0–5 impact scale with predefined behavioral anchors for each score level (as detailed in Additional file 1). Each of the 18 cases was then independently categorized into one of three primary pitfall categories by the two radiologists. Initial disagreements in categorization, which occurred in 3 of the 18 cases (16.7%), were resolved through a consensus discussion involving a third senior thoracic radiologist (>20 years of experience) to reach a final definitive classification. Statistical analysis Statistical analyses were performed using R software (version 4.5.1; R Foundation for Statistical Computing, Vienna, Austria). Continuous variables are reported as mean ± SD, and categorical variables as frequencies and percentages. Exact 95% confidence intervals (CI) for binomial proportions were calculated using the Clopper-Pearson method. The inter-rater reliability of the initial, independent pitfall categorizations was assessed using Cohen's kappa (κ) coefficient and its 95% CI. A κ value of 0.61–0.80 was interpreted as substantial agreement. This study was exploratory in nature, and no a priori sample size calculation was performed. The primary goal was to develop and test the feasibility of a novel classification framework rather than to determine precise error rates. The final sample of 18 non-concordant cases was deemed sufficient for this qualitative purpose and proved adequate to demonstrate substantial inter-rater reliability (κ=0.68), supporting the framework's reliability for initial validation. Results Overall diagnostic performance The patient selection process is detailed in Fig. 1. The final study population comprised 20 patients (mean age, 55.1 ± 16.9 years; 11 women [55.0%]). Detailed characteristics are provided in Table 1. It must be emphasized that this cohort was intentionally enriched for diagnostic ambiguity through the exclusion of straightforward cases; consequently, the discordance metrics reported below reflect the characteristics of this selected sample and should not be interpreted as representative of overall institutional diagnostic performance. In this difficulty-enriched cohort, major or moderate diagnostic discordance was identified in 18 cases (90.0%; 95% CI: 68.3%, 98.8%).Table 1Demographic, clinical, and diagnostic characteristics of the 20 study patients.#Age (y)SexClinical contextInitial radiological impressionFinal diagnosis (Reference Standard)ConcordancePitfall category138FAsymptomatic; prior ectopic pregnancy surgeryNeurogenic tumorFibrotic nodule with calcificationIncorrectPitfall 1254FAsymptomatic; HTN (3 yr)Benign-appearing lesionSchwannomaPartialPitfall 1325FAsymptomatic; prior pulmonary TBBenign lesion (fibroma not excluded)Solitary fibrous tumorPartialPitfall 3464MPalpitations, chest tightness; HTN (10 yr)Suspected neurogenic tumorBronchogenic cystIncorrectPitfall 2546FAsymptomaticBenign-appearing lesionSchwannomaPartialPitfall 1662FCough, expectoration; HTN (3 yr); pulmonary TBSuspected pleural mesotheliomaSchwannomaIncorrectPitfall 1759FFatigue, cough, fever; severe anemiaSuspected EMHExtramedullary hematopoiesisCorrect-868FLocal tenderness; HTN (40 yr); breast noduleInflammatory or fibrous massBreast cancer metastasisIncorrectPitfall 3964MRight-sided chest painSuspected TBSquamous cell carcinomaIncorrectPitfall 11082FRight-sided chest pain; HTN (40 yr)Fibrous vs. neurogenic tumorPleomorphic liposarcomaIncorrectPitfall 11161FProgressive dyspnea; prior pleural mesotheliomaHydropneumothoraxPleural mesothelioma (Stage IV)IncorrectPitfall 21250MChest wall erythema, pain; prior GI perforationSuspected chest wall abscessChest wall TB abscessCorrect-1353MIncidental finding; hepatitis BAnterior mediastinal irregular soft-tissue massInvasive thymomaIncorrectPitfall 11467MProgressive chest pain; long-term smokerSuspected metastasisPrimary chondrosarcomaIncorrectPitfall 11572MPainless chest wall massSuspected lipomaWD liposarcomaIncorrectPitfall 11632MPersistent chest wall thickeningNon-specific soft-tissue thickeningDesmoid-type fibromatosisIncorrectPitfall 11734FAnterior chest wall erythema, edemaDiffuse internal mammary thickeningInflammatory breast carcinomaIncorrectPitfall 11885MAbdominal bloating, hematochezia; colon cancerNon-specific subcutaneous noduleColon cancer metastasisIncorrectPitfall 31929MChest wall swelling, feverAbscessTB cold abscessIncorrectPitfall 12057FShoulder and back painSuspected apical chest wall metastasisPancoast tumor with chest wall invasionPartialPitfall 2Concordance between the initial radiological impression and the final diagnosis was graded as follows: Correct, complete agreement; Partial, partially correct diagnosis with some diagnostic overlap but incomplete accuracy; Incorrect, discordant or major diagnostic error.A hyphen (-) indicates that information was not applicable or data was unavailable.Abbreviations: F, female; M, male; EMH, extramedullary hematopoiesis; GI, gastrointestinal; HTN, hypertension; TB, tuberculosis; WD, well-differentiated. Table 1Demographic, clinical, and diagnostic characteristics of the 20 study patients.#Age (y)SexClinical contextInitial radiological impressionFinal diagnosis (Reference Standard)ConcordancePitfall category138FAsymptomatic; prior ectopic pregnancy surgeryNeurogenic tumorFibrotic nodule with calcificationIncorrectPitfall 1254FAsymptomatic; HTN (3 yr)Benign-appearing lesionSchwannomaPartialPitfall 1325FAsymptomatic; prior pulmonary TBBenign lesion (fibroma not excluded)Solitary fibrous tumorPartialPitfall 3464MPalpitations, chest tightness; HTN (10 yr)Suspected neurogenic tumorBronchogenic cystIncorrectPitfall 2546FAsymptomaticBenign-appearing lesionSchwannomaPartialPitfall 1662FCough, expectoration; HTN (3 yr); pulmonary TBSuspected pleural mesotheliomaSchwannomaIncorrectPitfall 1759FFatigue, cough, fever; severe anemiaSuspected EMHExtramedullary hematopoiesisCorrect-868FLocal tenderness; HTN (40 yr); breast noduleInflammatory or fibrous massBreast cancer metastasisIncorrectPitfall 3964MRight-sided chest painSuspected TBSquamous cell carcinomaIncorrectPitfall 11082FRight-sided chest pain; HTN (40 yr)Fibrous vs. neurogenic tumorPleomorphic liposarcomaIncorrectPitfall 11161FProgressive dyspnea; prior pleural mesotheliomaHydropneumothoraxPleural mesothelioma (Stage IV)IncorrectPitfall 21250MChest wall erythema, pain; prior GI perforationSuspected chest wall abscessChest wall TB abscessCorrect-1353MIncidental finding; hepatitis BAnterior mediastinal irregular soft-tissue massInvasive thymomaIncorrectPitfall 11467MProgressive chest pain; long-term smokerSuspected metastasisPrimary chondrosarcomaIncorrectPitfall 11572MPainless chest wall massSuspected lipomaWD liposarcomaIncorrectPitfall 11632MPersistent chest wall thickeningNon-specific soft-tissue thickeningDesmoid-type fibromatosisIncorrectPitfall 11734FAnterior chest wall erythema, edemaDiffuse internal mammary thickeningInflammatory breast carcinomaIncorrectPitfall 11885MAbdominal bloating, hematochezia; colon cancerNon-specific subcutaneous noduleColon cancer metastasisIncorrectPitfall 31929MChest wall swelling, feverAbscessTB cold abscessIncorrectPitfall 12057FShoulder and back painSuspected apical chest wall metastasisPancoast tumor with chest wall invasionPartialPitfall 2Concordance between the initial radiological impression and the final diagnosis was graded as follows: Correct, complete agreement; Partial, partially correct diagnosis with some diagnostic overlap but incomplete accuracy; Incorrect, discordant or major diagnostic error.A hyphen (-) indicates that information was not applicable or data was unavailable.Abbreviations: F, female; M, male; EMH, extramedullary hematopoiesis; GI, gastrointestinal; HTN, hypertension; TB, tuberculosis; WD, well-differentiated. Demographic, clinical, and diagnostic characteristics of the 20 study patients. Concordance between the initial radiological impression and the final diagnosis was graded as follows: Correct, complete agreement; Partial, partially correct diagnosis with some diagnostic overlap but incomplete accuracy; Incorrect, discordant or major diagnostic error. A hyphen (-) indicates that information was not applicable or data was unavailable. Abbreviations: F, female; M, male; EMH, extramedullary hematopoiesis; GI, gastrointestinal; HTN, hypertension; TB, tuberculosis; WD, well-differentiated. The root cause analysis of the 18 non-concordant cases revealed that the errors clustered into three thematic pitfalls (Fig. 2). The inter-rater agreement for this categorization was substantial (Cohen's κ = 0.68; 95% CI: 0.41, 0.95). However, the wide confidence interval highlights the uncertainty inherent to the small sample size, suggesting the true agreement level could range from moderate to almost perfect. The detailed scoring matrix for the five RCA domains for each case is provided in the Additional file 2: Figure S1.Fig. 2Root cause analysis of the 18 non-concordant cases. This variable-width 100% stacked bar chart illustrates the proportional contribution of the four primary Root Cause Analysis (RCA) domains to each of the three main diagnostic pitfall categories. The height of each bar is proportional to the prevalence of that pitfall type (Pitfall 1, n = 12; Pitfall 2, n = 3; Pitfall 3, n = 3). Data labels within each segment indicate the absolute number of cases and their corresponding percentage contribution. The legend defines the RCA domains. Note that Technical or Systemic Factors (RCA Domain 5) were identified as secondary contributors in some cases but were not the primary cause of error in this cohort. Root cause analysis of the 18 non-concordant cases. This variable-width 100% stacked bar chart illustrates the proportional contribution of the four primary Root Cause Analysis (RCA) domains to each of the three main diagnostic pitfall categories. The height of each bar is proportional to the prevalence of that pitfall type (Pitfall 1, n = 12; Pitfall 2, n = 3; Pitfall 3, n = 3). Data labels within each segment indicate the absolute number of cases and their corresponding percentage contribution. The legend defines the RCA domains. Note that Technical or Systemic Factors (RCA Domain 5) were identified as secondary contributors in some cases but were not the primary cause of error in this cohort. Pitfall 1: Interpretive errors due to misleading features This was the most frequent error pattern, identified as the primary pitfall in a substantial majority of non-concordant cases (12/18, 66.7%; 95% CI: 41.0%, 86.7%). These errors were driven by nonspecific imaging features or limited awareness of rare entities. The archetypal example was Case #10 (Fig. 3a, b), where a large pleomorphic liposarcoma was misinterpreted as a benign fibrous tumor.Fig. 3Illustrative examples of the three primary diagnostic pitfalls. (a, b) Pitfall 1: Interpretive Error (Case #10). An unenhanced axial CT (a) shows a large, nonspecific soft-tissue mass (asterisk) in the right chest wall. The corresponding contrast-enhanced CT (b) reveals marked heterogeneous enhancement (arrow), inconsistent with a benign tumor. Final diagnosis was pleomorphic liposarcoma. (c, d) Pitfall 2: Perceptual Error (Case #11). Extensive nodular pleural thickening (arrows in c, mediastinal window) was overlooked due to a prominent hydropneumothorax, which dominated the initial review on the lung window (d). This satisfaction of search error led to the missed diagnosis of pleural mesothelioma. The overlooked pleural findings are also indicated by arrows in (d). (e, f) Pitfall 3: Error of Integration (Case #8). An enhancing nodule in the left chest wall (arrows on axial (e) and coronal (f) images) was mischaracterized as benign. This error resulted from the failure to integrate the critical clinical context of a known history of breast malignancy. Final diagnosis was breast cancer metastasis. Illustrative examples of the three primary diagnostic pitfalls. (a, b) Pitfall 1: Interpretive Error (Case #10). An unenhanced axial CT (a) shows a large, nonspecific soft-tissue mass (asterisk) in the right chest wall. The corresponding contrast-enhanced CT (b) reveals marked heterogeneous enhancement (arrow), inconsistent with a benign tumor. Final diagnosis was pleomorphic liposarcoma. (c, d) Pitfall 2: Perceptual Error (Case #11). Extensive nodular pleural thickening (arrows in c, mediastinal window) was overlooked due to a prominent hydropneumothorax, which dominated the initial review on the lung window (d). This satisfaction of search error led to the missed diagnosis of pleural mesothelioma. The overlooked pleural findings are also indicated by arrows in (d). (e, f) Pitfall 3: Error of Integration (Case #8). An enhancing nodule in the left chest wall (arrows on axial (e) and coronal (f) images) was mischaracterized as benign. This error resulted from the failure to integrate the critical clinical context of a known history of breast malignancy. Final diagnosis was breast cancer metastasis. Pitfall 2: Perceptual errors This pitfall, primarily driven by perceptual failures or cognitive biases, was identified in a smaller but significant subset of cases (3/18, 16.7%; 95% CI: 3.6%, 41.4%). In Case #11 (Fig. 3c, d), the initial diagnosis focused on a prominent hydropneumothorax, leading to the complete oversight of pathognomonic signs of mesothelioma. Pitfall 3: Errors of integration This error, characterized by the failure to synthesize imaging with patient history, was identified in an equal proportion of cases (3/18, 16.7%; 95% CI: 3.6%, 41.4%). Case #8 (Fig. 3e, f) demonstrated this pitfall, where a new, enhancing nodule in a patient with a known history of a breast malignancy was mischaracterized as a benign entity. Discussion In this hypothesis-generating, exploratory taxonomy-development study, we systematically analyzed the root causes of diagnostic error in CT interpretation of juxtapleural lesions and proposed a novel classification of three fundamental diagnostic pitfalls. This study was explicitly designed as a feasibility analysis of our novel RCA framework, not as a definitive assessment of error prevalence or clinical performance. While the sample size is modest (n = 20), the cohort was intentionally enriched with diagnostically challenging cases to maximize the identification of systematic failure mechanisms—a methodological strategy appropriate for taxonomy development but one that precludes generalization of error rates. This enrichment strategy follows the approach validated in the peer review literature. Itri et al. showed that nonrandom, targeted peer review identified 130 high-severity discrepancy cases (scores 3–4) from 190 reviewed cases, while random review of 1690 cases in the same practice found none [27]. Trinh et al. reported that a random program of 42,891 reviews produced meaningful discrepancies in only 0.5% of cases, with yield declining over time [28]. Random sampling simply cannot generate enough informative errors to support mechanism-level taxonomy development. It is crucial to emphasize that the high discordance rate (90.0%) observed is not presented as an indicator of general radiological performance, but rather as confirmation that our study sample was appropriate for the intended goal: to deeply analyze the taxonomy of errors in scenarios where they are most likely to occur. A key strength of our methodology was the structured RCA approach, which allowed us to move beyond simple error cataloging and achieve substantial inter-rater reliability for our pitfall taxonomy, providing preliminary evidence for the construct validity of the framework. The κ of 0.68 achieved here can be compared with published reliability data for conventional scoring systems. Bender et al. evaluated RADPEER using 21 subspecialty radiologists and reported κ of 0.11 (standard 4-point scale) and 0.20 (dichotomized), reflecting only slight-to-fair agreement [29]. Although the study designs differ, this contrast suggests that a mechanism-oriented taxonomy with behavioral anchors may be more reproducible than severity-based scoring. That said, κ of 0.68 still leaves roughly one-third of the beyond-chance variance unexplained, confirming that even structured RCA carries subjective elements that require iterative refinement and consensus-building. Our study's primary contribution is the proposed linkage between the underlying root causes and the observable clinical pitfalls. Unlike generic RCA models [12], our framework specifically maps thoracic imaging pitfalls to cognitive error mechanisms, enabling targeted debiasing strategies. This framework maps the five domains of our RCA model (as detailed in Additional file 1) to the three pitfall categories. Pitfall 1 (Interpretive Errors) is primarily driven by Nonspecific Imaging Findings (Domain 1) and Knowledge Gaps (Domain 2), representing failures of pattern recognition. Pitfall 2 (Perceptual Errors) directly corresponds to Perceptual and Cognitive Errors (Domain 3), such as satisfaction of search. Finally, Pitfall 3 (Errors of Integration) results from a failure in Inadequate Clinical Information Integration (Domain 4). Systemic Factors (Domain 5) act as latent conditions that can exacerbate all three pitfalls. The predominance of Pitfall 1 (66.7%) in our cohort is a critical finding, though it must be interpreted with caution given the small sample size (n = 12 for this subgroup). This observation, while preliminary and requiring confirmation in larger cohorts, may reflect the specific challenges encountered in a county-level hospital. Radiologists in such settings often face a higher workload and function as generalists with limited daily exposure to rare entities, which heightens dependence on conventional pattern recognition [17], [20]. Kasalak et al. have demonstrated a quantitative link between work overload and increased diagnostic errors, a pressure that is often amplified in non-tertiary centers. This finding highlights a potential role for teleradiology subspecialty consultations and AI-driven decision support tools in these settings, though the efficacy of such interventions remains to be demonstrated in prospective trials. The distribution of pitfall types is therefore likely context-dependent. In a tertiary academic center with subspecialist thoracic radiologists and routine multidisciplinary team (MDT) meetings, one might hypothesize a lower incidence of knowledge-based interpretive errors (Pitfall 1), potentially offset by a relative increase in subtle perceptual or complex integration errors driven by higher case complexity. However, this hypothesis awaits empirical testing through multicenter comparative studies. Conversely, the higher case complexity and referral bias in such centers might lead to a relative increase in the proportion of Perceptual (Pitfall 2) and Integration (Pitfall 3) errors, where even experts can be challenged by subtle findings or convoluted clinical histories, a trend suggested by Lee et al. [20]. Pitfall 2 highlights the profound impact of cognitive biases like “satisfaction of search” [10], [14], [15]. The failure to appreciate the circumferential spread of mesothelioma (Case 11) reinforces that a disciplined, systematic review of the entire imaging volume is paramount. We postulate that future AI tools could be designed not merely as lesion detectors, but as “cognitive debiasing” agents that could, for instance, automatically query the electronic medical record or prompt confirmation of specific ancillary signs, thereby mitigating premature closure [21], [22]. However, these proposals remain speculative; the development and clinical validation of such AI debiasing tools represent a substantial research agenda that is still in its infancy. Beyond technological solutions, mitigating perceptual errors requires systemic process changes. For example, implementing a mandatory ‘structured visual search’ checklist for complex thoracic cases could force a systematic review of all key anatomical areas (pleura, mediastinum, lung parenchyma), reducing the risk of satisfaction of search. The effectiveness of such checklists, however, requires empirical validation in radiology-specific settings. Pitfall 3 was the most clearly avoidable. The mischaracterization of metastatic disease (Case 8) represents a fundamental breakdown in data synthesis, often compounded by systemic barriers to information access, providing direct, case-based evidence supporting the European Society of Radiology’s (ESR) “Integrative Diagnostics” initiative, as outlined by Beauchamp et al. [23]. Our findings on Pitfall 3 provide a clear, real-world rationale for this initiative’s call to deeply integrate clinical information into the imaging report. This underscores the need to shift from an isolated, pattern-recognition model to a framework of “synthesized diagnostics” [24]. Cochon et al. used the SEIPS (Systems Engineering Initiative for Patient Safety) framework to study diagnostic errors in radiology across five data sources, finding that no single source captured the full error spectrum [30]. Integration-type failures are therefore probably undercounted when only retrospective image review is used, as in this study. Mitigating Pitfall 3 requires robust integration of clinical context at the point of care. Practical interventions could include developing ‘smart’ structured reporting templates where critical fields like ‘known malignancies’ or ‘relevant history’ are mandatory. Furthermore, emerging generative large language models (LLMs) hold conceptual promise for automatically summarizing unstructured clinical notes from the electronic medical record and presenting a concise, relevant clinical summary to the radiologist at the time of interpretation, a concept explored by Sun et al. [25]. It is important to note that this application remains an area of active research; the feasibility of LLM-based clinical summarization depends critically on EMR data quality, and its implementation must address the inherent risk of AI “hallucinations” through robust radiologist verification protocols that have yet to be validated in clinical practice. Mitigating these identified pitfalls requires a multi-faceted approach, including risk-stratification checklists, structured reporting, and MDT meetings [11], [26]. This exploratory study has several important limitations that should be considered when interpreting its findings. First and most critically, the deliberate enrichment of the cohort for diagnostically difficult cases represents a significant methodological trade-off: while this strategy was essential for maximizing the yield of informative errors for taxonomy development, it fundamentally precludes generalization of the observed discordance rate (90.0%) to routine clinical practice, where straightforward cases constitute the majority of examinations. The prevalence of each pitfall type in an unselected, consecutive cohort remains unknown and is likely substantially different. Second, the small sample size (n = 20, with only 18 non-concordant cases for analysis) limits statistical precision, as reflected by the wide confidence intervals throughout (e.g., Cohen’s κ 95% CI: 0.41–0.95; Pitfall 1 proportion 95% CI: 41.0%–86.7%), and precludes subgroup or predictor analyses. Cohen’s κ is also sensitive to uneven category distributions; with Pitfall 1 comprising 66.7% of errors, future work should report complementary metrics such as Gwet’s AC1 alongside κ [32]. Third, all cases were sourced from a single county-level hospital, which may limit the generalizability of our pitfall distribution to tertiary academic centers with different case mixes, subspecialty expertise, and systemic factors. Fourth, the retrospective design introduces inherent selection and information biases. Fifth, the study period spanned a transition between two different PACS systems (Jiangxi Zhongtong Bit and Mindray); while both are DICOM-compliant and reviewers were proficient in both, we cannot completely exclude subtle systemic influences on reading efficiency. Sixth, the RCA scoring process, while structured with predefined anchors and Delphi-based domain refinement, retains a degree of subjectivity, and potential consensus bias cannot be fully excluded as all reviewers were from affiliated institutions. Finally, the proposed taxonomy has not yet undergone external validation, broader reproducibility testing across different reader experience levels, or assessment of its impact on diagnostic performance, educational outcomes, or patient safety metrics. Looking forward, the framework presented here serves as a foundational step. The essential next stage is to validate this three-pitfall taxonomy in a prospective, multicenter study that includes tertiary academic hospitals, multiple independent readers with varying experience levels, and assessment of whether the framework can improve diagnostic accuracy when incorporated into educational curricula or clinical workflows. Such a study would not only test the generalizability of our findings but also allow for comparative analysis of pitfall distribution across different healthcare settings and identification of system-level contributors. Ultimately, a validated and refined framework could form the basis for developing targeted educational interventions and, potentially, automated decision support tools. However, the translation from taxonomy to clinical impact requires multiple rigorous validation steps that remain to be undertaken. The taxonomy could also support morbidity and mortality conferences as a classification tool, following the resident-oriented M&M format shown to be feasible by Taussig et al. [31], allowing error patterns to be tracked over time and teaching to target specific failure mechanisms. Conclusions In conclusion, this exploratory study proposes a novel three-pitfall taxonomy for classifying diagnostic errors in the CT interpretation of juxtapleural lesions. The framework demonstrated feasibility and achieved preliminary evidence of substantial inter-rater reliability (κ = 0.68), supporting its potential as a structured tool for error analysis. However, the findings are derived from a small, difficulty-enriched, single-center cohort and have not yet undergone external validation. The taxonomy should therefore be viewed as a conceptual foundation for future research rather than a validated instrument ready for clinical deployment. Prospective, multicenter studies with larger, unselected cohorts and independent readers are required to establish the framework’s generalizability, construct validity, and impact on diagnostic performance, education, and patient outcomes before it can be recommended for routine quality assurance or training programs. Ethics approval and consent to participate This retrospective observational study involving human participants was approved by the Institutional Review Board of Shicheng County People's Hospital (Reference number: ZZUIRB 2025–07). The need for written informed consent was waived by the same ethics committee due to the retrospective nature of the study and the use of fully anonymized data. The study was conducted in accordance with the Declaration of Helsinki. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. CRediT authorship contribution statement Honggang Xu: Writing – review & editing, Supervision, Methodology, Conceptualization. Suiqiao Huang: Writing – review & editing. Huixian Chen: Investigation, Funding acquisition, Data curation. Kunquan Lai: Writing – original draft, Investigation, Formal analysis, Data curation. Huan Lai: Investigation, Formal analysis, Data curation. Xinhuan Huang: Investigation, Formal analysis, Data curation. Wu yongshun: Writing – review & editing, Supervision, Project administration, Methodology, Conceptualization. Declaration of Competing Interest The authors declare that they have no competing interests.