Software-assisted structured reporting and semi-automated TNM classification for NSCLC staging in a multicenter proof of concept study

Abstract

Objectives: In this multi-center study, we proposed a structured reporting (SR) framework for non-small cell lung cancer (NSCLC) and developed a software-assisted tool to automatically translate image-based findings and annotations into TNM classifications. The aim of this study was to validate the software-assisted SR tool for NSCLC, assess its potential clinical impact in a proof-of-concept study, and evaluate current reporting standards in participating institutions.

Methods: A framework for SR and staging of NSCLC was developed in a multi-center collaboration. SR annotations and descriptions were used to generate semi-automated TNM classification. The SR and TNM classification tools were evaluated by nine radiologists on n = 20 representative [18F]FDG PET/CT studies and compared to the free text reporting (FTR) strategy. Results were compared to a multidisciplinary team reference using a generalized linear mixed model (GLMM). Additionally, participants were surveyed on their experience with SR and TNM classification.

Results: Overall, GLMM analysis revealed that readers using SR were 1.707 (CI: 1.137-2.585) times more likely to correctly classify TNM status compared to FTR strategy (p = 0.01) resulting in increased overall TNM correctness in 71.9% (128/178) of cases compared to 62.8% (113/180) FTR. The primary source of variation in classification accuracy was explained by case complexity. Participants rated the potential impact of SR and semi-automated TNM classification as positive across all categories with improved scores after template validation.

Conclusion: This multi-center study yielded an effective software-assisted SR framework for NSCLC. The SR and semi-automated classification tool improved TNM classification and were perceived as valuable.

Critical relevance statement: Software-assisted SR provides robust input for semi-automated rule-based TNM classification in non-small-cell lung carcinoma (NSCLC), improves TNM correctness compared to FTR, and was perceived as valuable by radiology physicians.

Key points: SR and TNM classification are underutilized across participating centers for NSCLC staging. Software-assisted SR has emerged as a promising strategy for oncologic assessment. Software-assisted SR facilitates semi-automated TNM classification with improved staging accuracy compared to free-text reports in NSCLC.

Keywords: Lung; Non-small-cell lung carcinoma; PET-CT; TNM classification.

Fig. 1

Participants' opinion on SR and perceived barriers of clinical implementation. The perception of physicians to general statements regarding SR and potential barriers regarding SR are shown in (a, b), respectively. The survey reflects an overall positive perception regarding SR and perspectives on its clinical implementation. Participants rated a lack of digital infrastructure and perceived increased reporting time as the most relevant obstacles to clinical translation. SR, structured reporting

Fig. 2

Classification performance (orange = incorrect and green = correct) of study participants with regard to individual TNM-descriptors, demonstrating improved accuracy of SR strategy across all categories compared to FTR. SR, structured reporting; FTR, free text reporting

Fig. 3

Representative classification errors in SR and FTR. A Shows the primary tumor in patient 3 with broad visceral pleural contact indicative of an infiltration of the visceral pleura (T2a), as compared to size-based stage T1c. B Shows the primary tumor in patient 13 demonstrating a maximum multiplanar diameter in the coronal plane of 4.9 cm (T2b) as compared to the maximum axial diameter of 3.9 cm (T2a). C Shows a right-sided axillary lymph node metastasis consistent with an extrathoracic metastasis instead of a regional lymph node (N3) as it is not included in the International Association for the Study of Lung Cancer (IASLC) map. SR, structured reporting; FTR, free text reporting; AC, attenuation correction; FDG, fluorodeoxyglucose

Fig. 4

Spider plot visualizing the differences in assessment of SR (n = 9), plotting was based on the 7-point Likert scale responses (ranging from −3, “strongly disagree,” to +3, “strongly agree”). Questions 22–36 provided in Table 1 are presented in clockwise order. Overall, SR was perceived as a valuable reporting strategy across all categories by participants. Significantly improved perceptions are highlighted with an asterisk (*), demonstrating superior performance of the evaluated SR and classification tool. SR, structured reporting