Presenter: Kaushik Dutta
Imaging Science 2nd year Ph.D. student
Dr. Kooresh I. Shoghi
Manual delineation by experts is considered a gold standard in segmentation of tumors in preclinical setting. However, manual method is time intensive and has high interobserver variability along with low reproducibility. Our work aims to develop and evaluate the performance of a convolutional neural network (CNN) algorithm for automatic localization and segmentation of triple negative breast cancer (TNBC) patient-derived tumor xenografts (PDX) tumors in preclinical T2 weighted and T1-weighted MR images. Advanced quantitative methods like radiomics facilitates the extraction of higher dimensional data from the radiological images. The imaging biomarkers derived from radiomic analysis in order to exhibit clinical relevance needs like predicting treatment response needs to be robust and reproducible. Our automated pipeline extracts high dimensional radiomics features from the segmented regions and analyze the sensitivity of the features to tumor boundaries along with analyzing the reliability and robustness of the deep learning-based segmentation to that of manual annotation.