Explainable Machine Learning for Breast Cancer Growth Forecasting using MRI Time Series

Evaluating growth of a breast cancer tumour is crucial in ensuring timely treatment for individual patients, however the majority of existing models rely on static data and do not take any temporal awareness into account. This paper proposes an Explainable Machine Learning pipeline for forecasting Future Tumour Volume in Breast Cancer Patients from MRI Time Series data. We trained temporal deep learning models over features of half a million lesion descriptors, including volume, shape irregularity, surface area and ADC values from 200 patients using Long Short Term Memory (LSTM), Temporal Convolutional Network (TCN) and Transformer. For example, the LSTM model performed best with RMSE of 2.86 cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup>, MAE of 2.03 cm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">3</sup> and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> score of 0.89 In order to interpretability, SHAP and LIME were used to interpret model predictions on a global basis as well as per patient based. The findings suggest the application of interpretable tumour growth prediction to pinpoint predictive features and clinical decision making support in breast cancer care, enhancing forecast accuracy and transparency.