This repository provides a supplementary code for Temporal Representation Learning of Phenotype Trajectories for pCR Prediction in Breast Cancer paper. In this work we designed a representation learning method for treatment/dissease progression learning.
Multi-task representation learning balances reconstruction performance (L_Rec) with temporal continuity of trajectories (L_Temp) and alignment of changes in responders (L_Align). A U-shaped denoising network extracts multi-scale features via its encoder. An MTAN-inspired masking module is used to steer attention across these tasks. The resulting trajectory representations are utilised for predicting pCR using a linear classifier. To see the integration of the losses please refer to /utils/ARTLoss.py.
The dataset used is a subset of 585 patients from the ISPY-2 cohort. Please see /data/data_splits.txt for the full list of patient IDs used. For each patient and each timepoint we use three DCE-derived images from ISPY-2 dataset:
- early enhancement (PE_early, 120–150 sec post-contrast)
- late enhancement (PE_late, ∼450 sec)
- signal enhancement ratio (SER = PE_early / PE_late)
To reduce memory usage, we generated axial-plane maximum intensity projections (MIPs) of the three DCE-derived volumes.
