Automatic Patient Eligibility for Photon-counting CT Using Discriminative and Generative AI Models in Neuroradiology

Area of Science:

• Artificial Intelligence in Medical Imaging
• Radiology Workflow Optimization
• Natural Language Processing for Clinical Decision Support

Background:

• Photon-Counting CT (PCCT) offers superior resolution and spectral data but generates large image volumes, straining radiologist and Picture Archiving and Communication Systems (PACS) capacity.
• Current neuroradiology lacks guidelines for PCCT versus conventional Energy Integrating Detector (EID) CT selection, making manual decision-making impractical.
• Natural Language Processing (NLP)-based large language models (LLMs) present a potential solution for automating the routing of CT requests to appropriate scanner technology.

Purpose of the Study:

• To evaluate the efficacy of discriminative and generative LLMs in automatically classifying neuroradiology CT requests for Photon-Counting CT (PCCT) versus Energy Integrating Detector (EID) CT.
• To determine if AI models can accurately identify requests necessitating advanced PCCT protocols, thereby optimizing scanner utilization.
• To compare the performance of various transformer-based discriminative models and general-purpose generative LLMs for this classification task.

Main Methods:

• A retrospective study analyzed 800 Spanish-language neuroradiology CT requests from a Radiology Information System (RIS).
• Requests were independently labeled by two neuroradiologists as Basic Protocol (BP) or Advanced Protocol (AP) requiring PCCT, with expert consensus as the gold standard.
• Transformer classifiers (Spanish BERT, RoBERTa, Llama-3.1-8B, Mistral-7B) were fine-tuned for binary classification, while ChatGPT 5.2 and Gemini 3 were used in a zero-shot setting.

Main Results:

• Discriminative transformer models demonstrated high performance, with RoBERTa achieving the best results (accuracy 0.925, F1-score 0.935).
• RoBERTa significantly outperformed Llama, Mistral, Gemini, and ChatGPT in classifying CT requests for PCCT suitability.
• Generative LLMs (ChatGPT, Gemini) yielded the poorest overall performance in this neuroradiology CT request classification task.

Conclusions:

• Discriminative transformer models, particularly RoBERTa, are highly effective for automatically identifying advanced PCCT-eligible neuroradiology CT requests.
• This AI-driven approach significantly outperforms general-purpose generative LLMs in accuracy and efficiency for CT request routing.
• The findings support the use of AI for optimizing PCCT utilization and streamlining neuroradiology workflows through automated eligibility routing.