utilizing prior knowledge. Next, to further reduce the possibility of overfitting, Jianan presented an unsupervised clustering algorithm for identifying imaging phenotypes of colorectal cancer lesions. Inspired by the observation that imaging phenotypes do not necessarily correlate with molecular subtypes, he hypothesized that different subtypes of tumors may have similar appearances. With an autoencoder-based Gaussian mixture model, it became possible to address the overlapping appearance problem while performing unsupervised clustering. Lastly, Jianan proposed an outcome prediction algorithm specifically designed for multifocal and metastatic cancer. It remains unclear how multiple tumors contribute to patient outcome, and most medical imaging-based prognostic models focus solely on the primary lesion. Jianan proposed a multiple-instance neural network for integrating information from all tumor lesions. His hypothesis that using all tumor lesions for outcome prediction improves prediction accuracy was empirically validated. This approach also enabled the analysis of individual lesion aggressiveness without labels, with implications for understanding disease mechanism and clinical decision making. Predictions based on the approaches proposed in Jianan’s thesis achieved better prognostic value when compared to existing clinical and imaging biomarkers. In the future, Jianan plans to validate his findings in larger datasets and different types of cancers. He will also explore the integration of other data modalities to develop biologically-relevant imaging biomarkers. 37 Jianan Chen Computer Vision News
RkJQdWJsaXNoZXIy NTc3NzU=