Aim: Artificial intelligence (AI)-based breast cancer grading may help to overcome perceived limitations of human assessment. Here, the potential value of AI grade was evaluated at the molecular level and in predicting patient outcome.
Methods and results: A supervised convolutional neural network (CNN) model was trained on images of 612 breast cancers from The Cancer Genome Atlas (TCGA). The test set, obtained from the Cooperative Human Tissue Network (CHTN), comprised 1058 cancers with corresponding survival data. Upon reversal, a CNN was trained from images of 1537 CHTN cancers and tested on 397 TCGA cancers. In TCGA, mRNA models were trained using AI grade and Nottingham grade (NG) as labels. Performance of mRNA models in predicting patient outcome was evaluated using data from 1807 cancers from the Molecular Taxonomy of Breast Cancer International Consortium (METABRIC) cohort. In selecting images for training, nucleolar prominence determined high- versus low-grade cancer cells. In CHTN, NG corresponded to significant survival stratification in stages 1, 2 and 3 cancers, while AI grade showed significance in stages 1 and 2 and borderline in stage 3 tumours. In METABRIC, the mRNA model trained from AI grade was not significantly different to the NG-based model. The gene which best described AI grade was TRIP13, a gene involved with mitotic spindle assembly.
Conclusion: An AI grade trained from the morphologically distinctive feature of nucleolar prominence could transmit significant patient outcome information across three independent patient cohorts. AI grade shows promise in gene discovery and for second opinions.
Keywords: TRIP13; LASSO; RNA; deep learning; gene; machine learning; nucleoli.
© 2021 John Wiley & Sons Ltd.