Can Artificial Intelligence applied to a pre-treatment 18F-FDG PET/CT scan be used to predict two-year disease free survival for patients with Oesophageal Cancer?

Abstract

Oesophageal cancer is one of the leading causes of cancer death in the. It is potentially curable with surgery but carries a significant risk of complication; the decision to treat is important for patients and clinicians. Several authors have explored the potential link between radiomic data in positron emission tomography / computed tomography (PET/CT) images and treatment response. We propose an artificial intelligence method to predict disease-free survival (DFS) from PET imaging for patients with upper gastro-intestinal (GI) adenocarcinoma using a larger patient cohort than has previously been described in the UK. Furthermore, we propose investigating the effect of the Block Sequential Regularized Expectation Maximum (BSREM or “Q.Clear”) BSREM image reconstruction algorithm on radiomic signatures and overall machine learning performance with 3 key questions: 1. Are radiomic feature(s) from pre-treatment PET imaging linked to DFS? 2. Can artificial intelligence predict DFS from pre-treatment PET imaging? 3. Does BSREM affect radiomic features and the ability to predict DFS? We retrospectively analysed the staging PET/CT images of 144 patients with upper GI tract adenocarcinomas who underwent curative surgical treatment. We analysed 58 radiomic, 3 clinical features and 2 reconstruction methods (OSEM vs BSREM). We compared 6 machine learning (ML) algorithms for predicting DFS up to 2 years post treatment. We found that larger, heterogeneously distributed tumours were associated with poorer DFS rates. Radiomic features related to grey-level run length matrix were robust to different image reconstructions but features evaluating local variations, such as grey-level co-occurrence matrix contrast, were susceptible to reconstruction method. Most ML algorithms tested did not produce sufficient accuracy for use clinically however, BSREM images with a logistic regression algorithm, provided the most clinically relevant results: an overall 75% accuracy predicting 70% of successful, and crucially, 83% of failed treatments. Radiomic signatures from the PET images for upper GI cancer patients can aid clinicians and patients in identifying where closer monitoring for recurrence is required after surgical treatment. BSREM remains a useful tool for image quality enhancement but caution is advised when interpreting radiomic signatures. BSREM images with a logistic regression algorithm showed initial promise for predicting 2-year DFS from the radiomic signature of the primary tumour however further work with larger, standardised cohorts is required to validate this.

Date of Award	31 Dec 2023
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Jill Tipping (Supervisor)