A preliminary study on fusing CT and pathology information to improve the performance of tumor response prediction

Early stage prediction of chemotherapy response is critically important in treatment evaluation and management for advanced stage ovarian cancer patients. However, in clinical practice, the currently available markers cannot achieve satisfactory accuracy. The purpose of this investigation is to propose a novel method which fuses both CT and pathology image information to improve the performance of the current clinical marker for tumor response prediction. Accordingly, a total of 782 radiomics features were quantified from CT images after excluding low-variance features, representing the tumors’ geometric, density and textual properties. Meanwhile, the pathological features were extracted by one foundation model (CONCH, Contrastive learning from Captions for Histopathology). All the extracted information is fused at feature or score levels. For the feature level fusion, all the extracted features were concatenated together as the input of the following prediction model, which is established based on support vector machine (SVM). Score level fusion strategy establishes two SVM models separately for CT and pathology features, and two corresponding scores are then fused together. A total of 41 patients were collected for the experiment, including 32 responders and 9 non-responders. The result shows that feature level fusion approach achieves an AUC of 0.750 and an accuracy of 0.732, which is higher than the performance of CT or pathology information only model. Meanwhile, the score level fusing approaches yields an AUC of 0.708 and accuracy of 0.659. Our study initially verified the feasibility of fusing CT/pathology information to improve the performance and robustness of the clinical marker for chemotherapy response prediction.