Personalised Preoperative (Neoadjuvant) Chemotherapy (NACT) to Optimize Curative Treatment in Breast Cancer

Background: No effective method is available to predict, prior to NACT, which recipients' tumour will achieve the best response - a pathological complete response (pCR). Accurate detection of tumour response is essential to optimise patients treatment options.

Objectives: (1) To develop a breast tumour segmentation tool for automatic tumour localisation and measurement in routinely used contrast enhanced MRI for the qualitative assessment of tumour response before and after NACT; (2) To identify a set of imaging features to add to our ongoing projects of better predictive tools for pCR, before NACT; (3) To accurately diagnose pCR which can replace unnecessary surgery after NACT.

Project Plan: Ground-truth tumour masks will be firstly generated using our previously developed semi-automatic tool. Then a deep learning model will be trained to achieve automatic image segmentation. The imaging features will be extracted using radiomics measurements and deep leaning methods.

Feature selection methods will be applied to select a small subset of these features for predicting tumour response using pCR and recurrence-free survival time (RFS) as the endpoints. A computational model will be trained using a unique local dataset (400 patients). The developed method will be validated on another dataset (150 patients) from an independent local clinical trial.

We will also use a public dataset (202 patients) to externally validate the generalisability and repeatability of the developed method. Timeline: Month 4: delivery of automatic tumour segmentation tool. Month 12: delivery of validated computational model for tumour response prediction.

Impact: With actively involved in on-going international NACT trials, we plan to extent our analysis to include over 1,000 subjects from other UK trial centres for confirmation of the preliminary findings, and eventually embed the developed pipeline into routine clinical settings. The developed image segmentation and feature learning framework is highly likely to be adapted to other diseases.