JMIR Publications

ABSTRACT

Background:

Artificial intelligence research in image-guided oncology has grown exponentially, yet how far the field has progressed from diagnostic assistance toward direct therapeutic execution has never been quantified. Existing bibliometric surveys categorize studies by technical architecture or clinical domain, metrics that track publication volume but not proximity to procedural deployment.

Objective:

We developed a hierarchical functional classification framework to map the global landscape of therapeutic AI development across five major oncological indications. Our two specific objectives were: (1) to classify publications by clinical output function along the diagnostic-to-therapeutic continuum, and (2) to quantify the translation gap using three complementary metrics, triangulated against trial and device registries.

Methods:

We extracted 29,277 Web of Science publications spanning five image-guided oncologic specialties (thyroid, breast, lung, prostate, and liver) published between January 2010 and April 2026. AI-related records were classified by clinical function using a three-stage protocol: keyword categorization, contextual scoring, and rule-based filtering. Inter-rater reliability, validated on 518 independently coded publications, yielded Cohen's κ of 0.92. Our framework distinguished Diagnosis AI (disease identification) from therapeutic AI, then further stratified therapeutic AI into Bridge-support AI (treatment planning, prognosis, patient selection) and True Treatment AI. True Treatment AI was defined by concurrent satisfaction of two criteria: ≥Level 2 on the Yang Surgical Autonomy Scale and ≥Stage 1 on the IDEAL Framework.

Results:

Of 16,937 AI-related publications identified, 14,277 (84.3%) were categorized as Diagnosis AI and only 2,660 (15.7%) as therapeutic AI. All therapeutic publications fell exclusively within the Bridge-support tier. None satisfied the dual-framework criteria for True Treatment AI, yielding a uniform penetration rate of 0.00% across all five oncological domains. This complete execution vacuum persisted despite an 11-fold variation in inter-domain treatment-to-diagnosis ratios. The finding held under threshold relaxation, sensitivity analyses, and independent triangulation against 3,491 ClinicalTrials.gov records and 1,430 FDA device listings.

Conclusions:

Each specialty should periodically profile its diagnostic-to-therapeutic translational progress. The uniform absence of True Treatment AI across 15 years and five domains indicates that this gap is structural rather than cumulative, rooted in methodological inheritance from diagnostic paradigms and in regulatory category mismatches. Closing this gap requires coordinated framework development across regulatory, research, and clinical communities, rather than incremental algorithmic improvements.