What does this research mean for the field?

Parameter-efficient fine-tuning of biomedical language models using structured clinical findings can generate coherent, physician-style PET-CT reports with high factual correctness and low hallucination risk. Novelty: ClaimNovelty.METHODOLOGICAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

What question did this study set out to answer?

To develop and evaluate a domain-specific language model pipeline that converts structured metabolic and anatomical findings into coherent, safe, physician-style PET-CT reports.

September 5, 2026Indian Journal of Nuclear MedicineOpen Access

Context-Grounded PET-CT Report Generation Using LoRA-Fine-Tuned BioMedLM with Physician Validation and Safety Evaluation

View Full Paper

Ask AI

Bookmark

Authors

DPDibya PrakashHomi Bhabha National Institute PKPramukh KulkarniHomi Bhabha National Institute VRVenkatesh RangarajanHomi Bhabha National Institute

Discussion

Member takes

Implication

Validation study demonstrates reliable PET-CT report generation from structured clinical context, indicating feasibility for AI reporting pipelines.

Key Points

To develop and evaluate a domain-specific language model pipeline that converts structured metabolic and anatomical findings into coherent, safe, physician-style PET-CT reports.
Assembled 836 section-level PET-CT reports across 4 anatomical regions plus whole-body scans from 170 patients (85 with breast cancer, 85 controls).
Fine-tuned BioMedLM (2.7B) using Low-Rank Adaptation (LoRA) with an objective penalizing hallucinations and rewarding negation preservation, conditioned on structured findings such as SUVmax and lesion laterality.
Evaluated generation via five-fold cross-validation, held-out test sets, natural language processing metrics, and a blinded review by two nuclear medicine physicians across 28 test cases.
BioMedLM-LoRA achieved a ROUGE-L of 0.545 and BERTScore-F1 of 0.911 on held-out test data, outperforming template and retrieval baselines (all p < 10^-10), with 1.000 SUVmax faithfulness.
Independent physician review revealed a factual correctness score of 4.59/5, completeness of 4.77/5, clinical usefulness of 4.12/5, and 0 unsafe hallucinations (one-sided 95% CI: 0–10.7%).
Physicians rated 55.4% of reports as clinically acceptable Grade A (53.6% for Physician 1, 57.1% for Physician 2) with moderate inter-rater agreement (κ = 0.61).

Cite This Study

Prakash et al. (2026) studied this question.

synapsesocial.com/papers/6a9bd4436b95aff0620ebc87 https://doi.org/10.25259/ijnm_104_2026

View Full Paper

Ask AI

Bookmark

Implication

Validation study demonstrates reliable PET-CT report generation from structured clinical context, indicating feasibility for AI reporting pipelines.

Context-Grounded PET-CT Report Generation Using LoRA-Fine-Tuned BioMedLM with Physician Validation and Safety Evaluation

Discussion

Member takes

Implication

Key Points

Cite This Study

Context-Grounded PET-CT Report Generation Using LoRA-Fine-Tuned BioMedLM with Physician Validation and Safety Evaluation

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider