Abstract
With the latest advancements in optical bioimaging, rich structural and functional information has been generated from biological samples, which calls for capable computational tools to identify patterns and uncover relationships between optical characteristics and various biomedical conditions. Constrained by the existing knowledge of the novel signals obtained by those bioimaging techniques, precise and accurate ground truth annotations can be difficult to obtain. Here we present a weakly supervised deep learning framework for optical signature discovery based on inexact and incomplete supervision. The framework consists of a multiple instance learning-based classifier for the identification of regions of interest in coarsely labeled images and model interpretation techniques for optical signature discovery. We applied this framework to investigate human breast cancer-related optical signatures based on virtual histopathology enabled by simultaneous label-free autofluorescence multiharmonic microscopy (SLAM), with the goal of exploring unconventional cancer-related optical signatures from normal-appearing breast tissues. The framework has achieved an average area under the curve (AUC) of 0.975 on the cancer diagnosis task. In addition to well-known cancer biomarkers, non-obvious cancer-related patterns were revealed by the framework, including NAD(P)H-rich extracellular vesicles observed in normal-appearing breast cancer tissue, which facilitate new insights into the tumor microenvironment and field cancerization. This framework can be further extended to diverse imaging modalities and optical signature discovery tasks.