Investigation of the potential of artificial intelligence in imaging-based assessment of resectability in ovarian cancer

Investigation of the Potential of Artificial Intelligence in Imaging-Based Assessment of Resectability in Ovarian Cancer Objective: This study aims to analyze and compare the medical data and radiological images of ovarian cancer patients treated in our clinic with the resectability predictions provided by an artificial intelligence (AI) program. The primary objective is to evaluate the AI system's ability to predict resectability based on preoperative imaging and compare these predictions with radiology reports and the treatments applied in clinical practice. Additionally, the study investigates the effectiveness and reliability of AI-based systems in assessing resectability and guiding treatment planning in ovarian cancer. Materials and Methods: This retrospective study included patients diagnosed with ovarian cancer who were treated or followed at the Department of Obstetrics and Gynecology, Akdeniz University Hospital, between January 1, 2015, and January 1, 2025. A total of 230 patients were initially identified; however, 46 were excluded due to the unavailability of preoperative imaging or pathology data. The final analysis included 184 patients. Data were processed through the AI program ChatGPT. For each patient, both radiological imaging and clinical features were evaluated to generate AI-based resectability probabilities, which were then compared with predictions derived from radiology reports. Furthermore, the treatments suggested by AI were compared with those actually applied, and the time interval between diagnosis and treatment was assessed to evaluate the contribution of AI to decision-making timelines. Results: The mean age at diagnosis among the 184 patients was 55.67 ± 11.70 years. Of these, 69.0% were postmenopausal, 23.4% were smokers, and the average BMI was 27.72 ± 4.83 kg/m². Hypertension was the most common comorbidity. The mean CA-125 level was 1305.34 ± 3495.64 U/mL. Histopathologically, serous carcinoma was the most common subtype (69%), followed by endometrioid (10.9%), clear cell (5.4%), and granulosa cell tumors (3.3%). Among the 127 serous carcinoma patients, 89.8% were high-grade. Primary cytoreductive surgery was performed in 61.4% of cases. A statistically significant 56 difference (p < 0.001) was found between the AI-predicted resectability based on imaging and that based on radiology reports, with a mean discrepancy of 13.29%. In 64.1% of cases, the AI prediction was higher, equal in 7.1%, and lower in 28.8%. The average absolute difference in resectability probability was 14.28% for higher AI predictions and 14.34% for lower predictions. When analyzed by imaging modality, the discrepancy was 13.20% for CT, 12.86% for MRI, and 14.17% for PET. Regarding treatment planning, AI-generated treatment suggestions matched the actual clinical decisions in 75.5% of cases. The remaining 24.5% showed discordance, predominantly due to AI suggesting neoadjuvant chemotherapy while surgery was performed instead. Among those who underwent surgery despite AI recommending chemotherapy, 52.5% failed to achieve maximal cytoreduction. Of the 71 patients who received neoadjuvant chemotherapy, 69 were non-resectable, and 2 were resectable but received chemotherapy due to comorbidities—both cases aligned with AI suggestions. The average time from diagnosis to treatment initiation was 35.17 days. Conclusion: Ovarian cancer is often diagnosed at an advanced stage and requires effective planning of both diagnosis and treatment. Accurate evaluation of surgical resectability is essential for improving survival outcomes. This study demonstrates that AI-based systems show promising potential in predicting resectability and assisting clinical decision-making in ovarian cancer. When integrated into clinical workflows, AI may enhance diagnostic precision and reduce time to treatment. However, the observed discrepancies between AI recommendations and real-world practices indicate that AI should currently serve only as a supportive tool rather than an autonomous decision-maker. Its outputs must be critically assessed and tailored by clinicians. Further large-scale, multicenter studies are needed before routine clinical implementation. Keywords: Artificial Intelligence, Ovarian Cancer, Resectability, Imaging Modalities, CT, MRI, PET, Gynecologic Oncology