A recent study has explored the role of deep learning as a tool for rapid mutational screening in melanoma.
Detecting mutational status
In melanoma, 50-60% of mutations found occur in the BRAF oncogene. With the development of targeted therapies, the ability to determine the mutational status of the BRAF gene has become critical in the management of melanomas. Current methods to detect mutational status include DNA molecular assays and rapid screening tests (e.g. real-time polymerase chain reaction). All of these options however require tumour tissue for analysis.
Researchers have recently explored image-based analysis as an alternative method for mutation detection. This method is suitable for when a tumour is either not available or is inadequate for direct testing. Many of these analyses involve the use of radiomics. Although analyses have expanded to histopathology with the advent of digitised whole slide images (WSI).
The field of pathomics aims to extract and quantitate features from high-resolution digitised WSI on a large scale. The purpose of this is to integrate these features with molecular signatures, to develop biomarkers and to predict clinical or treatment outcomes. Pathomics could potentially generate a vast amount of data. Therefore, machine learning algorithms provide a unique opportunity to link features from images with an in depth understanding of tumour biology. Researchers have shown that deep convolutional neural networks (CNNs) can predict the presence of actionable genetic mutations in solid tumours using histopathological images.
Deep learning analysis
In this study, published as a pre-print in Biorxiv, researchers applied a deep CNN to whole slide images for prediction of mutated BRAF in melanomas resected from 256 patients. Researchers applied deep learning techniques to histopathology images of formalin-fixed paraffin-embedded (FPPE) primary melanomas. They then confirmed, through saliency mapping and pathomics analysis, that the mutated BRAF genotype was associated with detectable and quantifiable nuclear differences. This aimed to provide a genotype-phenotype link in melanoma tumour cells.
The model was able to first select for tumour-rich areas (Area Under the Curve=0.96), accurately differentiating melanomas from benign tissues. It then predicted for the presence of mutated BRAF genotype. They found that BRAF mutated cells exhibited larger and rounder nuclei, which could be predicted through deep learning and pathomics.
Deep learning-based mutational predictions are unlikely to replace direct molecular testing in the near future. However, there is great promise for these approaches to provide mutational data rapidly. This is particularly the case in situations where tissue is sparse and costs are low. The team believe that this approach could be scaled up in research and clinical settings.