Characterization of artificial intelligence performance for lesion detection using synthetic lesions in PET imaging
Bibliographic record
Abstract
Abstract Background The use of artificial intelligence (AI) for lesion detection has witnessed increased interest and efforts in recent years. Meanwhile, task‐specific characterization and comparison of AI performance is lacking as supportive evidence prior to its implementation in the clinical setting. Purpose To evaluate the use of synthetic lesions in positron emission tomography (PET) and computed tomography (CT) to characterize the performance of lesion detection AI in terms of their limits of detection. Methods An image library was constructed containing 565 well‐characterized synthetic lesions in 114 reconstructed studies from 56 real, disease‐free PET/CT patient data. Using the Lesion Synthesis Toolbox (LST), lesions were manually defined in terms of location, size and intensity. These lesions were then synthesized, forward projected, and added to the raw patient PET data before reconstruction using the same methods used clinically. Lesions were also appended to the reconstructed CT images. This library was sent to two external research teams developing AI for lesion detection in fluorodeoxyglucose ( 18 F‐FDG) PET. AI reported lesions were compared to ground truth data and labelled as hit, miss, or false positive. Psychophysical response models were fitted to each AI's responses to characterize their performance. Results Both AI methods confirmed higher lesion detection rates with increased lesion size and contrast. One AI consistently outperformed the other in terms of number of reported lesions, sensitivity, and precision. The fitted psychophysical response model demonstrated both graphically and parametrically an ability of this model to detect smaller lesions for a given degree of reliability. For example, 10 mm diameter lesions could be detected with 90% sensitivity at 8:1 versus 16:1 lesion to background ratio for the two algorithms. Likewise, 3:1 contrast lesions could be detected with 90% sensitivity when lesion diameters were approximately 16 and 31 mm for each algorithm respectively. Compared to defined lesions parameters, the corresponding AI segmented lesions had lower contrast, consistent with partial volume effects in PET imaging, and also smaller size. Conclusion Synthetic lesions are a useful tool to characterize the performance of lesion detection by an observer. Visual and psychometric response models of lesion detection performance with respect to lesion characteristics are effective to objectively compare AI performance on the merit of limits of detection. These methods can be applied to objectively compare lesion detection performance with any alternative decision support tool including human and machine observers, display technologies, and image generation systems.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.014 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.000 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".