(P 178) PREDICTION OF PRIMARY IMATINIB-REFRACTORY PATIENTS WITH LOCALLY ADVANCED OR METASTATIC GASTROINTESTINAL STROMAL TUMOR (GIST) USING RADIOMICS AND ARTIFICIAL INTELLIGENCE (AI)

Objective: Although GIST genotyping may be useful for predicting imatinib-sensitivity, some GIST, especially those not harboring KIT or PDGFRA mutations, are theoretically associated with a broad spectrum of response to treatment. Therefore selecting patients with imatinib primary resistance remains challenging. Recent advancements in the field of radiomics, which extract detailed quantitative features from medical imaging, show promising potential in identifying predictive markers for drug response. The integration of AI into these analyses is rapidly evolving. AI's capacity to process and analyze extensive and complex datasets enables the discovery of subtle patterns and correlations that may not be apparent through traditional methods. This emerging synergy between radiomic feature extraction and AI-driven analysis holds significant promise for research.

In the current project we aim to develop an AI model capable of predicting the response to imatinib based on detailed radiological assessments and advanced radiomics analyses, using CT images and clinical features of a retrospectively selected group of patients diagnosed with locally advanced or metastatic GIST, with various molecular background and different sensitivity to imatinib.

Methods: We retrospectively analyzed a total of 47 patients with locally advanced or metastatic GIST, of which 16 patients were included in the study because they fulfilled the inclusion criteria: (i) histological diagnosis of locally advanced or metastatic GIST, (ii) age older than 18 years, (iii) patients treated with imatinib as first line treatment and (iv) among the scans performed, availability of CT scans in DICOM format both prior to and one year following treatment. Expert radiologists manually segmented CT scans and extracted the volume of interest (VOI) for each patient, composed by principal mass and metastasis. Radiomics and more classical image features were extracted from VOIs. We performed a Leave-One-Out Validation procedure, considering as test the subject left out from the validation and training. Several machine learning models were trained to classify between: partial response, stable disease, and progressive disease following RECIST criteria v1.1 (Figure 1).

Results: The study retrospectively analyzed 16 patients diagnosed with GIST, with a mean age of 56 years and a gender distribution of 56% female and 44% male. Tumor localization was primarily gastric (56%), with ileal and duodenal locations each comprising 19%. Comprehensive molecular profiling at the time of diagnostic biopsy or during surgery identified ten patients with c-KIT exon 11 mutations (62%), one with c-KIT exon 9 mutation, and two cases with no KIT/PDGFRA mutation, while three patients had unknown molecular status. All patients received 400 mg daily imatinib without experiencing toxicity that required discontinuation. Among the scans performed, CT scans taken at baseline and one year post-treatment initiation were considered for assessing treatment response through a comparative analysis. The median progression-free survival was 50 months, ranging from 5 to 129 months. After one year 43.7% had stable disease, 37.5% had a partial response, and 18.8% showed disease progression according to RECIST criteria v1.1. Patients were identified in two groups: imatinib-sensitive (patients with partial response or stable disease) and imatinib-resistant (patients with progressive disease). The development of a radiomics signature, aimed at differentiating these groups, is currently ongoing. Detailed findings and the clinical implications of this signature will be disclosed at the upcoming meeting.

Conclusion: The integration of radiomics and artificial intelligence with clinicopathological data could offer a powerful approach to identifying patients at higher risk of imatinib resistance as early as the baseline CT scan. This technological synergy enables clinicians to enhance patient follow-up and detect relapses at an early stage, thereby significantly improving the effectiveness of therapeutic strategies.