Abstract Background Accurate non-invasive diagnosis of early-stage ovarian cancer remains challenging because of the limited number of biomarkers. Although artificial intelligence algorithms show promise for ovarian cancer diagnosis, their reliance on specialized engineering knowledge hinders their accessibility. The recent emergence of visual large language models such as GPT-4o has further expanded the potential of AI in this domain. Methods GPT-4o was trained to automatically recognize ovarian lesions, report key computed tomography (CT) features of ovarian lesions, and make a benign or malignant diagnosis based on these features. Radiologists and gynecologic oncologists independently reviewed the GPT-4o reports and evaluated GPT-4o's performance. Results GPT-4o achieved diagnostic accuracies of 80.80%, 79.14%, and 93.33% in the three datasets. Its performance surpassed that of gynecologic oncologist with 10 years of experience but was inferior to that of gynecologic oncologist with 16 years of experience and radiologists with ≥ 7 years of experience. The clinician-rated reliability in detecting the four key CT features was 4.22/5.00 for cyst wall and septum status; 4.24/5.00 for nodular or papillary protrusions; 4.30/5.00 for density and enhancement distribution; and 4.25/5.00 for cystic-solid characteristics. The use of GPT-4o increased the accuracy of radiologist and gynecologic oncologist diagnoses by 1.96% and 10.50%, respectively. Conclusions GPT-4o identifies the key CT features of ovarian cancer and achieves promising diagnostic accuracy with high-quality diagnostic evidence.
Zhang et al. (Tue,) studied this question.