The Lifecycle of a Neural Network in the Wild: A Multiple Instance Learning Study on Cancer Detection from Breast Biopsies Imaged with Novel Technique

Abstract

In the context of tissue examination for breast cancer assessment, we propose a label-free imaging based on Optical Coherence Tomography (OCT) signal combined with a multiple instance learning (MIL) model to respond to a critical need for fast at point-of-care diagnosis: biopsy or surgery time. This new imaging, Dynamic Cell Imaging (DCI), is the time-resolved variant of Full-Field OCT (FFOCT) and offers an intra-cellular resolution of about 1 micron, together with optical sectioning and an improved cell contrast. In order to tackle the challenges of limited data and annotations, while remaining in the scope of interpretability, we design an instance-level MIL model with a focus on adapted data sampling. The interest of this method is that it incorporates taskspecific feature learning and also produces instance predictions. For a dataset of 150 core-needle biopsies, we achieve a considerable improvement of more than 20 percentage points in specificity and about 10 in accuracy by leveraging intradomain (as compared to extra-domain) pre-training.