The SCIT program hosts a quarterly colloquium during which two trainees present the status of their research.

NEXT SEMINAR:

Wednesday, May 1, 2024
10:00 AM - 11:00 AM
In-Person (Lucas P083) and via ZOOM

PRESENTERS:

"3D Printed Microneedles for Breast Cancer Biomarker Discovery in Dermal Interstitial Fluid"
Andy Hung, MD
ABSTRACT: Dermal interstitial fluid (ISF) is an overlooked compartment rich in diagnostic biomarkers. Burgeoning interest in ISF diagnostics stems from the potential for anyone to self-collect ISF with minimal pain. ISF can act as a surrogate for serum or as a source of new biomarkers. Previously, we identified candidate ISF biomarkers in breast cancer. However, progress is hindered by the lack of a reliable method to quickly collect ISF in sufficient volumes. I present ongoing work in 3D printed microneedles designed to reliably collect at least 10 uL of ISF with minimal pain in under 5 minutes. 

"Prostate Cancer Detection Models: The Challenges of Adapting to Multi-Vendor Data"
Sara Saunders, PhD
ABSTRACT: Prostate cancer is the most common and second deadliest cancer among men in the US, but early detection can improve outcomes. Magnetic Resonance Imaging (MRI) can be used to guide biopsy procedures. However, the interpretation of MRI is challenging due to the subtle differences between cancer and normal tissue. Deep learning methods can assist in localizing lesions that can be targeted with biopsy or local treatment. However, deep learning models trained on one dataset often generalize poorly to new unseen datasets. For example, often deep learning models trained using MRIs from one vendor show poor generalization on MRIs acquired from different vendors. We aim to develop models that generalize well to new datasets, and the initial component of this is analyzing limitations in this aspect of current models.  We analyze difference in performance across vendors by training 2D and 3D convolutional neural networks known as the nnUNet, to detect biopsy confirmed clinically significant prostate cancer with 3450 multi-institutional, multi-vendor cases both acquired at Stanford and from the Prostate Imaging: Cancer AI (PICAI) dataset. We combined these 2D and 3D nnUNets into a single ensemble that outputs a map indicating the probability of prostate cancer at the voxel-level using multiparametric MRI. Evaluation of this ensemble on 642 test cases shows differences in performance across vendors of over 10% in Average Precision and over 7% in Area Under the Receiving Operator Curve, though the model was trained on data from all vendors. Analysis of these differences is a critical first step to the development of models that address the challenges of adapting to multi-vendor data.