Surgical planning for orbitotomy using artificial intelligence (ChatGPT)

Theme: Oculoplastic & reconstructive surgery
What: Oculoplastic & reconstructive surgery
Part of: Oculoplastics II: Cutting edge research in oculoplastics! / Oculoplastie II: Recherche de pointe en oculoplastie!
When: 5/31/2024, 02:00 PM - 03:30 PM
Where: Room | Salle 714 A

Abstract

Purpose: There has been an increasing use of artificial intelligence (AI) to aid radiologic imaging interpretation and to complement clinical decision making. ChatGPT is a recently developed large language model, and its ability to suggest appropriate radiologic imaging modalities for certain clinical presentations and to answer common patient questions has been investigated. However, its role in ophthalmologic surgical decision making has not been assessed. This study aims to assess the ability of ChatGPT to interpret diagnostic imaging reports and to recommend appropriate surgical approaches for patients undergoing orbitotomy.

Study design: We conducted a consecutive, retrospective case series of all adult orbitotomy cases from July 2021 to September 2023 of three oculoplastic surgeons at University of Toronto, Ontario, Canada.

Methods: Thirty-four patients underwent an orbitotomy. For each patient, the computed tomography (CT) or magnetic resonance imaging (MRI) report was input into ChatGPT 3.5. A standardized script was used each time to ask ChatGPT four questions: 1) Top 3 differential diagnosis; 2) Single most likely diagnosis; 3) Most appropriate type of biopsy (incisional vs. excisional) for the most likely diagnosis; and 4) Recommended surgical approach to access the lesion during an orbitotomy. For the recommended surgical approach, ChatGPT was given several multiple choice options based on the quadrant of the lesion. Our outcomes included the proportion of cases where ChatGPT’s differential diagnosis included the final pathology diagnosis and how often its recommended biopsy type and surgical approach matched the surgeon’s operative choice.

Results: The analysis included 30 eyes. The top 3 differential diagnoses proposed by ChatGPT based on the CT or MRI imaging report findings included the final pathology diagnosis in 50% of cases. When asked to single out the one most likely diagnosis, it matched the pathology diagnosis in 38% of cases. The suggested type of biopsy (incisional vs. excisional) matched the surgeon’s choice in 72% of cases. However, when asked regarding the most appropriate surgical approach to access the lesion in an orbitotomy, ChatGPT’s recommendation agreed with the surgeon’s choice in 39% of cases. For 5 patients, ChatGPT indicated neither type of biopsy as appropriate as the suspected diagnosis was inflammatory in nature. In two patients, it indicated neither biopsy was appropriate and suggested different biopsy techniques such as a stereotactic biopsy or fine-needle aspiration biopsy instead.

Conclusions: AI and ChatGPT demonstrate potential in diagnostic imaging interpretation and in aiding preoperative surgical planning. However there remains limitations in its ability to accurately interpret radiologic imaging findings without clinical context and to select an appropriate surgical approach, illustrating the complexity and nuance of this decision.

Presenter(s)

Presenting Author: Jenny Ma

Additional Author(s):

Kenneth Chang, Department of Ophthalmology and Vision Sciences, University of Toronto, Toronto, ON, Canada

Georges Nassrallah, Department of Ophthalmology and Vision Sciences, University of Toronto, Toronto, ON, Canada

---