Using GPT‑4o with CT exams helps diagnose ovarian cancer earlier

An AI approach that uses GPT‑4o to analyze pelvic CT exams could help clinicians diagnose early‑stage ovarian cancer more accurately and consistently, according to a study published February 17 in the Annals of Surgical Oncology.

"[Our research found that] GPT-4o identifies the key CT features of ovarian cancer and achieves promising diagnostic accuracy with high-quality diagnostic evidence," wrote a team led by Shimin Zhang, MD, of Shengjing Hospital of China Medical University in Liaoning, China.

Early detection of ovarian cancer can be a challenge, and more than half of cases are still diagnosed at metastatic stages, contributing to a five‑year survival rate of 31.4% compared with more than 90% when the disease is confined to the ovaries, the group explained. The gold standard for ovarian cancer diagnosis has long been surgical pathology, and clinicians rely heavily on imaging -- particularly pelvic CT -- to guide preoperative assessment.

But CT interpretation depends on the radiologist's experience and can be affected by high interobserver variability. Zhang's group explored the use of GPT-4o to improve CT's ability to diagnose early ovarian cancer via a study that included 479 patients with pathologically confirmed benign or early‑stage malignant ovarian lesions. The researchers trained the algorithm to identify four CT features associated with malignancy -- cyst wall and septum characteristics, nodular or papillary protrusions, density and enhancement patterns, and cystic versus solid composition -- and to recognize ovarian lesions, report key CT features of ovarian lesions, and make a benign or malignant diagnosis based on these features.

They found the following:

• GPT-4o achieved diagnostic accuracies of 80.8%, 79.1%, and 93.3% in the three datasets, respectively (i.e., ovarian lesions, key CT features of ovarian lesions, and benign or malignant diagnoses).
   
• Clinician ratings for GPT‑4o's reliability in detecting four key CT features were 4.2 for cyst wall and septum status; 4.2 for nodular or papillary protrusions; 4.3 for density and enhancement distribution; and 4.2 for cystic-solid characteristics (all scores out of 5).
   
• GPT-4o increased the accuracy of radiologist and gynecologic oncologist diagnoses by 1.96% and 10.5%, respectively.

The group noted that using GPT-4o with pelvic CT images improved the diagnostic performance of less experienced clinicians, reporting that the diagnostic accuracy of gynecologic oncologists with less than seven years of experience increased from 67.9% to 78.1% when they were assisted by the model.

"With further validation in diverse datasets, GPT-4o holds promise as a novel approach for early ovarian cancer detection, ultimately improving the current landscape of early-stage ovarian cancer management," the authors concluded.

Access the full study here.