Implementing machine learning methods to integrate radiological and pathological data to assess treatment response in oesophageal/gastro-oesophageal cancer

Lead Supervisor
Dr Anita Grigoriadis
Reader in Cancer Bioinformatics
School of Cancer and Pharmaceutical Science, FoLSMs
King’s College London
sabine.landau@kcl.ac.uk

Co-supervisor
Professor Vicky Goh, Professor Gary Cook, Dr Kasia Owczarczyk

Project Details

Background:
Oesophageal/gastro-oesophageal cancer is a cancer of unmet need with less than 50% of patients surviving 5 years even after curative treatment (1). Predictive models based on clinical variables and visual assessment of standard-of-care imaging alone cannot reliably predict chemotherapy response; a poor response occurring in up to 60% of patients in whom alternative treatment strategies may be more effective (2). Novel radiotherapy-immunotherapy combinations hold promise as means of treatment intensification in gastro-oesophageal cancer, however, the benefit appears to be confined to a subgroup of patients (3).
There is a clinical need for better predictive tools to identify non-responders pre-operatively and gain understanding of the underlying biology to allow selective treatment escalation.
Integration of multiparametric MRI imaging and digital pathology images, using advanced image post-processing methods (radiopathomics), has resulted in more accurate patient classification and outcome prediction compared to either of these imaging modalities alone e.g. in brain tumours (4).
In our previous work, MRI radiomic biomarkers in addition to clinical variables demonstrated higher predictive performance for treatment response in oesophageal cancer (5). We have also shown that machine learning using a convolutional neural network of 18F-Fluorodeoxyglucose PET data could improve neoadjuvant chemotherapy response prediction in oesophageal cancer (6).

This PhD project will combine medical images (radiomics -MRI- and pathomics – digital pathological images) with clinical data to develop new ways of predicting response to neoadjuvant chemotherapy in oesophageal cancers (Fig 1). The student will develop computational approaches to explore histopathology and MRI images to uncover features which in the hands of the treating physicians, pathologists, and radiologists, would add clinically relevant guidance to their current diagnostic information whether a patient would respond to a particular therapy.

Work Package 1: The student will implement and test several deep learning methodologies to explore digitised whole slide images of paired diagnostic biopsies and surgical resections. The aim is to identify features associated with pathological response assessed according to Mandard Tumour Regression Grade, as well as classifiers to predict response to chemotherapy prior to surgery allowing for treatment intensification, especially with radiotherapy-immunotherapy combination. Digitised pathological whole slide images will inform of intra-tumoural heterogeneity, the presence of intra-tumoural, and peri-tumoural immune infiltrates of the primary tumour.

Work Package 2: Radiomic features, such as wavelength, intensity, shape, and texture, can complement and extend pathological tumour and peri-tumoural characteristics. As a second objective, the student will work on convolution neural network to bridge clinically relevant information coming from MRI data with patient-matched and tumour-matched digitised pathological whole slide images. Depending on the MRI characteristics, classification, regression and clustering approaches will be evaluated.

Work Package 3: Working towards translation into clinical practice, developed approaches for predicting response to neoadjuvant therapy will be further tested in an external test dataset to validate performance and compared with current clinical practice and published radiomic approaches. Associations between selected MRI & histopathological variables with disease-free and overall survival will also be explored.

This PhD will enable the student to develop excellence in complex medical informatics research, including supervised and unsupervised machine learning techniques, methods to handle data collinearity and overfitting, classification of high dimensional, mixed data, model interpretation and clinical outcome prediction as well as experience in advanced image analysis, image post-processing, radiomic feature extraction and radiopathomic data integration. The student will benefit from multi-disciplinary and outcome analyses as well as academic radiologists, pathologists, oncologists and clinical data managers. The student will further have access to the expertise and resources of the London Medical Imaging and AI Centre for Value Based Healthcare, and will participate in weekly group meetings and monthly departmental seminars.

References:

  1. https://www.cancerresearchuk.org/health-professional/cancer-statistics/statistics-by-cancer-type/oesophageal-cancer
  2. Knight WRC, Zylstra J, Wulaningsih W, et al: Impact of incremental circumferential resection margin distance on overall survival and recurrence in oesophageal adenocarcinoma. BJS Open 2:229-237, 2018
  3. Kelly RJ, Ajani JA, Kuzdzal J, et al: Adjuvant Nivolumab in Resected Esophageal or Gastroesophageal Junction Cancer. N Engl J Med 384:1191-1203, 2021
  4. Chiu FY, Le NQK, Chen CY. A Multiparametric MRI-Based Radiomics Analysis to Efficiently Classify Tumor Subregions of Glioblastoma: A Pilot Study in Machine Learning. J Clin Med. 2021; 10(9):2030. Published 2021 May 10. doi:10.3390/jcm10092030
  5. Owczarczyk K WS, Grzeda M , Yip C ,Qureshi A ,Gossage J ,Davies A ,Cook G ,Goh V: Exploratory magnetic resonance imaging histogram biomarkers for response prediction to neoadjuvant treatment in oesophageal/gastro-oesophageal cancer. Annals of Oncology (2018) 32:S193-S194, 2021
  6. Ypsilantis PP, Siddique M, Sohn HM, et al: Predicting Response to Neoadjuvant Chemotherapy with PET Imaging Using Convolutional Neural Networks. PLoS One 10:e0137036, 2015

Datasets

The main dataset to be used in this project is the MIMOSA (Multiparametric MRI in Oesophageal Cancer Assessment) clinical trial (IRAS 115752). All patients participating in this study have given their informed consent to future research use of collected tissue samples as well as imaging and clinical data. An amendment has been submitted to Health Research Agency (HRA) to include the proposed analysis as an additional secondary endpoint of this study.

Keywords

Machine learning, radiomics, digital pathology, oesophageal cancer