Eugene Leung

Objective:Accurate interpretation of planar ventilation-perfusion (V/Q) scintigraphy, used for diagnosing pulmonary embolism (PE) based on PIOPED/EANM guidelines, requires objective assessment of mismatched V/Q defects. Manual delineation of V/Q defects is time-consuming, subject to interobserver variability, and rarely performed in practice, limiting standardized reporting and quantification of disease burden. To address these challenges, we evaluated four modern AI models for automated segmentation of vascular perfusion defects in planar V/Q scans and compared their performance to human annotators.Approach:We retrospectively identified 2,118 patients who underwent planar V/Q scans at The Ottawa Hospital (June 2019-February 2023). Six standard projections (ANT, POST, LAO, RAO, LPO, RPO) were included. Four 2D neural networks (U-Net, nnU-Net, Swin UNETR, and a Bottleneck Transformer U-Net [BTU-Net]) were trained on 1,313 patients (7,878 projections) and validated on 329 (1,974 projections) using physician-annotated defects. A hold-out test set of 46 high probability patients was used to evaluate segmentation quality, and defect detection accuracy using free-response receiver operating characteristic (FROC) analysis.Main results:BTU-Net was the only model performing on par with human readers, showing robust sensitivity across the entire range of segmentation probabilities. At 1.5 false positives per projection rate (FPPR), BTU-Net outperformed other models with sensitivity of 0.529 ± 0.026, On a separate hold-out set of low likelihood of disease patients (n=430), the lowest FPPR was 0.08 ± 0.01 for BTU-Net (P<0.0001).Significance:BTU-Net enables rapid, consistent, and accurate interpretation of planar V/Q scans. Such tools may enhance diagnostic efficiency, standardize reporting, and support non-expert readers in evaluating PE.

Background: Accurate reporting in nuclear medicine is essential for clinical decision-making. Trainees often generate preliminary reports with variable quality, and artificial intelligence (AI) tools such as ChatGPT-4o may enhance report clarity and accuracy, particularly in the impression section of the report. Objective: This study aimed to evaluate and compare the quality of positron emission tomography / computed tomography report impression sections generated by trainees and by the AI chatbot ChatGPT-4o (OpenAI), focusing on correctness, clarity, completeness, organization, use of diagnostic certainty terminology, and physician satisfaction. Methods: The impression sections of 200 positron emission tomography/ computed tomography reports generated by trainees and AI (100 reports each) were compared. The AI generated the impressions based on the stem, clinical history, and technical sections of the trainee-generated reports. Reports were blindly rated by 3 nuclear medicine physicians. Survey questions, including Likert-scale questions, assessed correctness, use of certainty terms, clarity, completeness, organization, and satisfaction. Statistical analyses included Fisher exact test, 2-tailed t tests, chi-square tests, ANOVA, and effect sizes (Cohen d ). Thematic analysis was performed on free-text comments. Results: AI-generated impressions were rated as correct in 92% of reports (mean 0.92, SD 0.27) and trainee-generated impressions rated as correct in 91% of reports (mean 0.91, SD 0.29; P =.80, negligible effect size). AI-generated impressions included certainty terms more frequently than trainee-generated impressions (94% vs 81%, respectively; P =.005; χ ² ₂ =6.58, small effect size). AI also used higher-level certainty terms more frequently than trainees (mean 4.23, SD 1.25 vs mean 3.69, SD 1.93 on a 5-point scale; P =.02). Clarity was high in both groups (AI: mean 4.49, SD 0.70, and 89.8% of reports rated as clear vs trainees: mean 4.35, SD 0.85, and 87.0% of reports rated as clear; P =.20). Completeness was significantly higher for AI-generated impressions than trainee-generated impressions (mean 4.52, SD 0.76, and 90.4% complete vs mean 4.09, SD 1.01, and 81.8% complete, respectively; P <.001, small-to-medium effect size). Organization was similar between the groups (mean 4.22, SD 0.91 vs mean 4.39, SD 0.74; P =.15). Satisfaction ratings were also compared, with AI-generated impressions achieving a mean score of 4.31 (SD 0.77) and 86.2% satisfaction, compared with a mean score of 4.14 (SD 0.91) and 82.8% satisfaction for trainee-generated impressions ( P =.16). Thematic analysis showed that trainee-generated impressions were more frequently criticized for accuracy (74% vs 19%; P <.001) and actionability (20% vs 4.8%; P =.02), whereas AI-generated impressions were more frequently criticized for excessive length (33% vs 0%; P <.001). Conclusions: These findings suggest that the AI chatbot ChatGPT-4o can serve as a valuable adjunct in nuclear medicine reporting, particularly at the resident level, by improving decisiveness and completeness while complementing trainee education and clinical oversight.

BACKGROUND Accurate reporting in nuclear medicine is essential for clinical decision- making. Trainees often generate preliminary reports with variable quality, and AI tools such as ChatGPT may enhance report clarity and accuracy particularly in the impression section of the report. OBJECTIVE To evaluate and compare the quality of PET/CT report impression sections generated by trainees and by the AI chatbot ChatGPT, focusing on correctness, clarity, completeness, organization, use of diagnostic certainty terminology, and physician satisfaction. METHODS We compared the impression section of 200 PET-CT reports, which was generated by trainees and AI (n=100) each. The AI-generated the impression based on the stem, clinical history and technical section of the trainees generated reports. Reports were blindly rated by three nuclear medicine physicians. A survey of questions including Likert-scale questions assessed correctness, certainty word use, clarity, completeness, organization, and satisfaction. Statistical analyses included Fisher’s test, t-tests, chi- square, ANOVA, and effect sizes (Cohen’s d). Thematic analysis was performed on free- text comments. RESULTS - Correctness: Comparable between groups (AI: 92% correct, mean=0.92 ±0.27; trainees: 91% correct, mean=0.91 ±0.29; p=0.80, negligible effect). - Certainty word use: AI included certainty terms in 94% of reports versus 81% for trainees (p=0.005, χ²=6.58, small effect size). - Type of certainty word: AI used higher-level certainty terms more frequently (mean=4.23 vs 3.69 on 5-point scale; p=0.02). - Clarity: High in both groups (AI 4.49 ±0.70, 89.8% rated clear vs trainees 4.35 ±0.85, 87%; p=0.20). - Completeness: AI significantly outperformed trainees (AI 4.52 ±0.76, 90.4% complete vs trainees 4.09 ±1.01, 81.8%; p<0.001, small–medium effect). - Organization: Similar (AI 4.22 vs 4.39; p=0.15, not significant). - Satisfaction: Similar (AI 4.31 ±0.77, 86.2% satisfied vs trainees 4.14 ±0.91, 82.8%; p=0.16). Thematic analysis showed trainees were more often criticized for Accuracy (74% vs 19%, p<0.001) and Actionability (20% vs 4.8%, p=0.02), while AI was flagged for long impression (33% vs 0%, p<0.001). CONCLUSIONS These findings suggest that AI chatbot, Chat GPT can serve as a valuable adjunct in nuclear medicine reporting, particularly at the resident level, by improving decisiveness and completeness while complementing trainee education and clinical oversight. CLINICALTRIAL Not applicable. This study did not involve a prospective clinical trial.