Integration of Biomechanics and Machine Learning for Extracellular Vesicle Analysis

Presenter

Isabella Rasku-Casas

Group Members

Teniola Ogundeko

Campus

UMass Boston

Sponsor

Joanna Dahl, Department of Engineering, UMass Boston

Schedule

Session 3, 1:30 PM - 2:15 PM [Schedule by Time][Poster Grid for Time/Location]

Location

Poster Board A55, Campus Center Auditorium, Row 3 (A41-A60) [Poster Location Map]

Abstract

Addressing the challenge of slow manual image analysis in cancer-related applications, we aim to enhance precision in identifying and analyzing intricate cellular behavior, challenging the limitations imposed by human assessment. Our focus lies in developing advanced image processing technology to improve object detection in extracellular vesicles (EVs). EVs are cell-released particles containing vital molecular information and are crucial in understanding cellular dynamics. 

Our methodology involves optimizing image formats and meticulously labeling EVs using polygonal masks within a dataset of 500 samples. Initially opting for YOLO, a traditional object detection model, we recognized its limitations in handling diverse EV shapes and background blending tendencies. Consequently, we transitioned to leveraging Mask R-CNN's advanced object detection capabilities, significantly enhancing precision, especially in identifying tiny, faint EVs in high-magnification (100x) microscopy images. 

Following training, our model achieved an impressive mean average precision (mAP) of 84%. mAP evaluates both precision (accuracy of positive predictions) and recall (ability to detect all positive instances), reflecting high accuracy in identifying and delineating EVs. This choice underscores the limitations inherent in traditional models like YOLO, emphasizing Mask R-CNN's efficacy in refining diagnostic processes. 

Refining EV detection, mapping its trajectories, and analyzing shapes using Mask R-CNN, our research optimizes the analysis pipeline. This enhancement ensures a more efficient and accurate evaluation, reducing human variability and expediting the analysis time. Our research enables a broader scope of studies to further explore the technology's potential in enhancing cancer diagnostics and treatment. 

Keywords

Image Analysis, Mask R-CNN, Object Detection, Diagnostic Technology, Machine Learning

Research Area

Cancer Studies

SIMILAR ABSTRACTS (BY KEYWORD)