Author: Bowen Jing, Jing Wang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose: Medical images acquired at multiple time points during neoadjuvant chemotherapy allow physicians to assess patients’ responses and personalize treatment plans accordingly. Studies from the I-SPY 2 clinical trial show that dynamic contrast-enhanced magnetic resonance (DCEMR) imaging is useful in predicting pathological complete response (pCR) for each patient. However, previous prediction methods usually rely on images from late time points (12th week or post-treatment) to ensure accuracy, which may outweigh the benefits of a response-based adaptive treatment strategy, as patients may have to undergo too many cycles of ineffective chemotherapy. To improve prediction accuracy at the early stage of treatment (3rd week), we propose a two-stage dual-task deep learning strategy for early pCR prediction using early time point data only.
Methods: Our proposed strategy was developed and evaluated on the I-SPY 2 dataset, which includes DCEMR images acquired at three time points: pretreatment (T0), after 3 weeks of treatment (T1), and after 12 weeks of treatment (T2). In the first stage, a network was trained to extract the latent-space image representation at T2. In the second stage, a dual-task convolutional long short-term memory network (LSTM) network was trained to predict pCR and the image representation at T2 using images from T0 and T1. This trained network enabled the prediction of pCR earlier without using images from T2.
Results: The conventional single-stage single-task strategy achieved an area under the receiver operating characteristic curve (AUROC) of 0.799. By using the proposed two-stage dual-task learning strategy, the AUROC significantly improved to 0.820 (p<0.01, DeLong’s test).
Conclusion: The two-stage dual-task learning strategy can significantly improve model performance for predicting pCR at the early time point (3rd week) of neoadjuvant chemotherapy. This early prediction model can potentially help physicians intervene earlier and develop personalized treatment plans accordingly.