PRISM - Machine Learning for Discovery of Pre-neoplastic signature in Mesothelioma

Background: Malignant pleural mesothelioma (MM) is a highly aggressive cancer and has a five-year survival rate of less than 10% from diagnosis.

The prognosis of patients with MM is poor since most of the patients are diagnosed at an advanced stage when MM is resistant to current treatment options MM is mainly caused by exposure to asbestos. Mesotheliomas arise most commonly from the mesothelial cells that line the pleura.

To diagnose MM, a biopsy is taken, which can be either in the form of pleural tissue (for histopathology) or pleural fluid (for cytology) or both.

By identifying early cancer changes and preneoplastic changes, at risk patients can be entered into a surveillance programme allowing earlier diagnosis of MM.

There is a strong clinical priority to develop novel systems that can analyse and detect pre-neoplastic changes in mesothelial cells for early cancer detection, improved patient stratification and preneoplastic biomarker discovery.

Aims: The aim of this proposal is to use machine learning techniques to identify phenotype markers of pre-malignant change in mesothelial cells. This will create a basis for tool development for early cancer detection in mesothelioma.

Methods: A data lake will be constructed containing clinical data, follow up data and Whole Slide Images from approximately 2050 MM cases, 360 atypical mesothelial proliferation and 3000 cases with normal or reactive mesothelial samples with WSI using H&E sections of biopsy and cytology samples available from the Royal Brompton and Harefield Hospitals NHS Trust (RBHT), Basildon and Thurrock Hospitals NHS Trust (BTHT), MESO bank (14), St George’s Hospital University of London (SGUL) and the Medizinische Hochschule Hannover (MHH).

Recent advances in deep learning neural networks have led to the possibility to integrate heterogeneous feature data, to label unlabelled data and to classify with noisy labels. All these techniques will be investigated, enhanced and applied in this project.

Deep learning techniques will applied such as data integration using heterogeneous feature space learning for discovery of MPM signatures, Data labelling for unlabelled data employing semi-supervised or transductive learning and learning with abstention. Outcome: This will create a basis for tool development for early cancer detection in mesothelioma.

Features extracted from MM can be used as markers for surveillance of pre-malignant changes in mesothelial cells and subsequently aid identification of early stage cancers when treatments can be more effective.

This will be the first machine learning method to identify pre-malignant changes in mesothelioma from digital pathological images.