A Machine Learning Framework for Oncological Margin Detection
Machine Learning , Breast Cancer , Margin Detection , Self-supervised Learning , Autoencoder , Multi-Instance Learning , weak labels , Basal Cell Carcinoma , iKnife , Mass Spectrometry
Cancer is the cause of one in four deaths in North America. A common phase of early-stage oncological treatment is the surgical resection of cancerous tissue. The presence of cancer cells on the resection margin, referred to as positive margin, is correlated with the recurrence of cancer and demands re-operation. The iKnife is a mass spectrometry modality that produces real time margin information based on the signatures of metabolites in surgical smoke. Using this modality and real-time tissue classification, surgeons could remove all the cancerous tissue during the initial surgery, improving patient survival, mental health, and cosmetic outcomes. An obstacle in developing iKnife cancer recognition models is the destructive, time consuming and sensitive nature of the data collection that limits the size and quality of the datasets. We propose to use machine learning techniques to create robust and generalizable cancer recognition models for the iKnife. To do so, we first perform a feasibility study using an easily accessible cancer type, basal cell carcinoma (BCC). This high incidence, low metastases cancer allows us to capture a larger dataset. We follow this work by implementing an attention based multi-instance learning model (MIL), which can better compensate for our small dataset by training on groups of samples given one collective label (e.g. cancer or benign) to learn the signatures of BCC. Finally, we propose to address the target dataset of breast cancer by building a self-supervised learning model from limited, weakly-labeled data. This model also explores the applications of domain transfer by first, learning to contextualize the general features of iKnife data using the BCC dataset. Following this pretraining, the model is then applied to our target breast cancer classification task. Through extensive experiments, we show the benefits of using machine learning frameworks for the development of iKnife oncological recognition models. Throughout our studies, we also focus on exploring the biochemical significance of the data by uncovering patterns and investigating biochemical attentions that support current literature. In the future, this work can be extended to other tissue types and translated to intraoperative use for real-time prediction of tissue signatures at the tumor margin.