Evaluating Human-AI Interaction in Pediatric Whole-Body MRI: An Exploratory Study of an AI-Assisted Tumour Overlay (Preprint)
Bibliographic record
Abstract
BACKGROUND: Artificial intelligence (AI) tools have the potential to enhance personalized clinical care, particularly in radiology. However, their integration into clinical workflows remains complex, especially in pediatric oncology, where early cancer detection is critical. Children with Li-Fraumeni Syndrome (LFS), a rare cancer predisposition disorder, undergo regular surveillance whole-body MRI (wbMRI), which presents an opportunity for AI-assisted tumour detection. OBJECTIVE: We evaluated the feasibility of an AI-assisted overlay for highlighting tumour-like regions in pediatric surveillance wbMRI and explored how access to the overlay influenced radiologist workflow, candidate-lesion marking behaviour, follow-up recommendations, and perceived workload. METHODS: We developed a patch-based AI segmentation model trained on augmented 2D slices from 675 surveillance wbMRI volumes of pediatric patients with LFS. The model was designed to highlight regions with high tumour probability. A reader study was conducted with two radiologists who independently reviewed wbMRI cases both with and without AI assistance. We measured evaluation time, number and location of tumours identified, type of follow-up recommendation, and subjective feedback using structured questionnaires. RESULTS: AI assistance altered interpretation workflows for both radiologists, with mixed effects. On average, the time required to evaluate each case increased when using the AI tool for both radiologists. However, one radiologist had an increase in the number of candidate lesion locations selected with the tool, and one had a decrease in the number of candidate lesion locations selected with the tool. Subjective feedback indicated that one of the radiologists found a greater difference in their perception between performing with the tool versus without the AI tool; however, both radiologists felt that the task was less difficult and less stressful with the AI tool. Inter-rater variability was evident, underscoring the need for personalized calibration of AI tools. CONCLUSIONS: AI-assisted wbMRI interpretation can improve tumour detection in pediatric cancer surveillance by reducing false negatives. However, its influence on workflow efficiency and inter-radiologist variability highlights the importance of careful implementation. Successful integration requires addressing challenges such as improving the predictive precision of AI models, offering intuitive end-user designs and instructions, and building trust in AI outputs. AI outputs can influence workflow and behaviour in reader-specific ways. Clinical translation will require larger, randomized, multi-reader studies and model refinement to reduce false positives and quantify lesion-level reader performance. This can help ensure better patient outcomes in addition to reduced clinician burnout.
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 | 0.002 | 0.013 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.000 |
| Insufficient payload (model declined to judge) | 0.002 | 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".