Acquisition Time-Informed Breast Tumor Segmentation from Dynamic Contrast-Enhanced MRI

TL;DR

This work tackles the challenge of variable tissue appearance across DCE-MRI acquisition times by conditioning segmentation networks on continuous acquisition times via Feature-wise Linear Modulation (FiLM). A FiLM generator produces per-channel scaling and shifting coefficients from the input times, and these modulate intermediate features throughout the encoder and decoder, enabling time-aware segmentation without stacking all phases as inputs. On the MAMA-MIA multisite dataset and an external Yunnan dataset, acquisition-time FiLM improves segmentation accuracy and robustness, with the largest gains when modulation is applied across the full network; transformer-based backbones show data-dependent behavior. The approach offers a practical, low-overhead method to exploit DCE-MRI dynamics for more reliable tumor delineation in breast cancer.

Abstract

Dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) plays an important role in breast cancer screening, tumor assessment, and treatment planning and monitoring. The dynamic changes in contrast in different tissues help to highlight the tumor in post-contrast images. However, varying acquisition protocols and individual factors result in large variation in the appearance of tissues, even for images acquired in the same phase (e.g., first post-contrast phase), making automated tumor segmentation challenging. Here, we propose a tumor segmentation method that leverages knowledge of the image acquisition time to modulate model features according to the specific acquisition sequence. We incorporate the acquisition times using feature-wise linear modulation (FiLM) layers, a lightweight method for incorporating temporal information that also allows for capitalizing on the full, variables number of images acquired per imaging study. We trained baseline and different configurations for the time-modulated models with varying backbone architectures on a large public multisite breast DCE-MRI dataset. Evaluation on in-domain images and a public out-of-domain dataset showed that incorporating knowledge of phase acquisition time improved tumor segmentation performance and model generalization.

Figures (3)

• Figure 1: Visualization of DCE-MRI enhancement. From left to right are pre-contrast, first post-contrast, second post-contrast, and expert segmentation of the tumor overlay on the first post-contrast image.
• Figure 2: Proposed Architecture. A common segmentation architecture often contains an encoder and a decoder block as well as skip connections. We modulate intermediate feature maps in the encoder and decoder stages. We experimented with four different configurations.
• Figure 3: Qualitative comparison of tumor segmentation across FiLM configurations. DCE-MRI slice showing predictions from the baseline nnU-Net (Vanilla) and four FiLM placement strategies (Encoder, Decoder, Bottleneck, and Encoder–Decoder modulation). Baseline model yields most false positives without infused temporal enrichment information. The best-performing configuration (Encoder–Decoder Modulated) yields much fewer false positives. Ground truth (GT) is shown in the bottom right.