Deep Learning-Based Classification for Circulating Tumor Cells

Circulating tumor cells are a key biomarker in liquid biopsy, offering a noninvasive approach to monitor and guide therapeutic decision in cancer patients. The extreme rarity and the heterogeneity of CTCs make their identification extremely challenging. To overcome these limitations, CTC can be isolated exploiting an unbiased workflow based on Parsortix® strategy, which enriches for CTC according to their size and deformability, in combination with DEPArray ${ }^{\text {TM }}$ technology that allows visualization and selection of single CTC. Then, expert physicians manually analyze the digital images of cells acquired with DEPArray technology. The introduction of an automatic system for CTC identification into clinical workflows could facilitate the early detection of metastases, optimize treatment decision-making, and improve patient outcomes. In this paper, we present a Deep Learning-based classification pipeline to distinguish CTCs from leukocytes in a liquid biopsy, focusing on improving diagnostic procedure and accuracy. The proposed method automatically classifies images acquired with DEParray technology, obtained at “Fondazione IRCCS-Istituto Nazionale dei Tumori” of Milan. It is based on ResNet architecture, a Convolutional Neural Network for image analysis and very popular in the medical field. The introduction of a carefully designed augmentation process allows incrementing data variability, obtaining improving in the model performance. The proposed method achieves an F1 score of $0,777 \pm 0,012$. The obtained promising results encourage us to do future research in this field.