Abstract PD6-04: Tumor Boundary and Shape Features are Predictive of Recurrence in Triple Negative Breast Cancer (TNBC) Patients

Abstract Background: TNBC patients exhibit an aggressive disease course with at least 25% developing recurrence or metastasis within 3-5 years, despite standard-of-care therapies. Clinicopathologic factors (CP; age, growth pattern, tumor size, margin status and grade) have limited value in identifying high risk TNBC patients. Early and accurate prediction of recurrence in TNBC patients from clinical mammograms would facilitate therapy optimization. However, new strategies to identify high-risk patients are needed. In this study, we used radiomics features at the tumor boundary to predict recurrence in TNBC patients and LLM/GPT to explain the prediction in natural language. Materials and methods: Mammograms from node negative TNBC patients (n=77; aged 30-90 years, lesion size of 2-45 mm, grades 2 and 3) who underwent adjuvant chemotherapy and had 5-yr follow-up were analyzed. Lesions were manually segmented and tumor boundary (1 mm containing both intra-tumoral and peritumoral regions) was automatically determined. Over 2,000 radiomics features from the tumor boundary and the central regions of the tumor were extracted to quantify heterogeneity, texture, shape, and size of the tumor. Recursive feature elimination using non-linear random forest classifier was used to reduce feature dimensionality and prevent over-fitting. A random forest classifier was employed to build an AI/machine learning model for risk of recurrence using the reduced feature dimension. Top ranking features along with clinicopathological variables were further used in a Bayesian network (BN) to create an interpretable machine learning model. To generate a patient-specific report, a sentence transformer model was further used for BN graph embedding, which aided in querying the graph in natural language inheriting the node/feature specific dependencies of the BN in LLM/GPT (BN LLM). Results: Recurrence prediction was significantly improved by analysis of radiomics features at the tumor boundary (using continuous gradient magnitude and texture features) and shape features, as compared to the same features extracted from the entire tumor. In a 3-fold validation framework, gradient and texture features at tumor boundary, along with tumor shape, better predicted recurrence (mean AUC of 0.78 (std 0.15)) as compared to the same features from the entire tumor (mean AUC of 0.35 (std 0.11)). The addition of clinical variables did not improve the AUC. Using clinical variables alone, AUC of 0.6 (std 0.09) was obtained. Using BN LLM an example patient specific report in natural language is summarized below: Patient A clinical profile : Tumor size – large; age – younger (