Introduction
The incidence of oral cancer is increasing in the world, in particular developing countries, where the abuse of nicotine is rising. Approximately 350,000 new cases of oral cancer occur annually around the world. Using a cytobrush to obtain a couple of thousand cells extracted from the abnormal area, typically located in the mucosa or gingiva, can be a cheap, non-invasive technique for evaluating cells repetitively. The cells are analyzed at both the morphological level, through visualization with standard Papanicolaou staining, and at the protein level, by staining with antibodies that target specific structures of interest. Intracellular signal transduction pathways such as JAK-STAT and mTOR/Akt are correlated to cell differentiation and cell survival, which are often upregulated in cancer tissues(Harsha. We aim to develop a machine learning pipeline to quantify and differentiate single cells based on signal intensity and provide a wider analysis of potential malignant cells.

Material and methods
Keratinocytes obtained from patients by cytobrush were placed on glass slides. 6 antibodies targeting intracellular proteins from JAK-STAT and mTOR/Akt pathways were added to the keratinocytes, and the in situ proximity ligation assay (isPLA) technique was applied. If two proteins are interacting, their specific antibodies carry oligonucleotides allowing a padlock probe to hybridize and ligate to a circle, and the circle templates a rolling circle amplification reaction (RCA). The RCA product was visualized by fluorophores hybridized to the complementary sequence. The signals were detected and imaged in fluorescence microscopy. An image analysis pipeline was developed in Python to segment individual cells and detect antibody signals using Cellpose and Big-FISH, respectively. The distances between fluorophores, nuclei, and cell borders were computed to quantify and characterize protein localization patterns. 

Results
A panel of 6 proteins correlated to JAk-STAT and mTOr/Akt pathway were visualized on keratinocytes.  The image analysis pipeline successfully segmented individual cells and quantified the fluorophore signals. On average, approximately 30 cells were identified per image. Based on the spatial distribution of the fluorophore signals, the interactions appeared to be localized mainly within the cytoplasm, predominantly near the cell periphery.
 
Discussion
Using a couple of thousands of cells collected by a cytobrush or other non-invasive techniques, one may find dysplastic cells early in the biological procedure or in patients that which suffer a relapse. We observed that the cell smear preparation is essential to facilitate the image analysis to segment the cells correctly.  Using cytology samples to study both morphology and molecular profile could potentially complement the conventional invasive histopathology approach, where the much less invasive cytological brush sampling allows more frequent sampling and shorter total turnaround time. Our image analysis pipeline has been adapted to handle the material wherein the keratinocytes are not homogeneously distributed and do not lay completely flat on the glass slide. Based on the identified interactions, it is in our immediate future work to investigate whether and in which ways these interactions differ between malignant and non-malignant cells.