Artificial intelligence for breast cancer precision pathology

Abstract

Breast cancer is the most common cancer type in women globally but is associated with a continuous decline in mortality rates. The improved prognosis can be partially attributed to effective treatments developed for subgroups of patients. However, nowadays, it remains challenging to optimise treatment plans for each individual. To improve disease outcome and to decrease the burden associated with unnecessary treatment and adverse drug effects, the current thesis aimed to develop artificial intelligence based tools to improve individualised medicine for breast cancer patients.

In study I, we developed a deep learning based model (DeepGrade) to stratify patients that were associated with intermediate risks. The model was optimised with haematoxylin and eosin (HE) stained whole slide images (WSIs) with grade 1 and 3 tumours and applied to stratify grade 2 tumours into grade 1-like (DG2-low) and grade 3-like (DG2-high) subgroups. The efficacy of the DeepGrade model was validated using recurrence free survival where the dichotomised groups exhibited an adjusted hazard ratio (HR) of 2.94 (95% confidence interval [CI] 1.24-6.97, P = 0.015). The observation was further confirmed in the external test cohort with an adjusted HR of 1.91 (95% CI: 1.11-3.29, P = 0.019).

In study II, we investigated whether deep learning models were capable of predicting gene expression levels using the morphological patterns from tumours. We optimised convolutional neural networks (CNNs) to predict mRNA expression for 17,695 genes using HE stained WSIs from the training set. An initial evaluation on the validation set showed that a significant correlation between the RNA-seq measurements and model predictions was observed for 52.75% of the genes. The models were further tested in the internal and external test sets. Besides, we compared the model's efficacy in predicting RNA-seq based proliferation scores. Lastly, the ability of capturing spatial gene expression variations for the optimised CNNs was evaluated and confirmed using spatial transcriptomics profiling.

In study III, we investigated the relationship between intra-tumour gene expression heterogeneity and patient survival outcomes. Deep learning models optimised from study II were applied to generate spatial gene expression predictions for the PAM50 gene panel. A set of 11 texture based features and one slide average gene expression feature per gene were extracted as input to train a Cox proportional hazards regression model with elastic net regularisation to predict patient risk of recurrence. Through nested cross-validation, the model dichotomised the training cohort into low and high risk groups with an adjusted HR of 2.1 (95% CI: 1.30-3.30, P = 0.002). The model was further validated on two external cohorts.

In study IV, we investigated the agreement between the Stratipath Breast, which is the modified, commercialised DeepGrade model developed in study I, and the Prosigna® test. Both tests sought to stratify patients with distinct prognosis. The outputs from Stratipath Breast comprise a risk score and a two-level risk stratification whereas the outputs from Prosigna® include the risk of recurrence score and a three-tier risk stratification. By comparing the number of patients assigned to ‘low’ or ‘high’ risk groups, we found an overall moderate agreement (76.09%) between the two tests. Besides, the risk scores by two tests also revealed a good correlation (Spearman's rho = 0.59, P = 1.16E-08). In addition, a good correlation was observed between the risk score from each test and the Ki67 index. The comparison was also carried out in the subgroup of patients with grade 2 tumours where similar but slightly dropped correlations were found.