Computer Vision for Detecting and Measuring Multicellular Tumor Shperoids of Prostate Cancer

Abstract

We present a deep learning model to apply computer vision to detect prostate cancer spheroid cultures and calculate their volume. Multicellular tumour spheroids, or simply spheroids, represent a three-dimensional in vitro model of cancer. Spheroids are being increasingly used in drug discovery due to their superior ability to mimic the tumor microenvironment compared to monolayer cell cultures. A reduction in spheroid size in response to treatment with anticancer agents is indicative of the success of the therapy. As such, accurate spheroid detection and volume estimation is critical in assays involving spheroids. Automating spheroid detection and measurement reduces manual labor, laboratory costs, and research time. Our system is implemented using Darkflow YOLOv2, a single-phase object detector, based on a twenty-four-layer convolutional neural network. The network is trained on the custom data of biochemically-generated spheroids and their corresponding images, which are then bound and detected with an F1-score of 76% and an IoU of 69%. Volume calculations applied to the identified spheroids resulted in a high volume estimation accuracy with only 3.99% average error.