TLDR: LesiOnTime is a novel AI model designed to improve the accuracy of small breast lesion segmentation in DCE-MRI scans. It achieves this by mimicking radiologists’ diagnostic workflows, jointly leveraging longitudinal imaging data (previous scans) through a Temporal Prior Attention (TPA) block and clinical BI-RADS scores via a BI-RADS Consistency Regularization (BCR) loss. This approach allows the model to dynamically integrate temporal information and embed clinical domain knowledge, leading to superior performance in early cancer detection compared to existing methods.
Breast cancer remains a significant health challenge globally, with early detection being crucial for improving survival rates. While traditional screening methods like mammography are effective, Dynamic Contrast-Enhanced MRI (DCE-MRI) is often recommended for high-risk individuals due to its superior sensitivity. However, accurately identifying small or subtle lesions in DCE-MRI scans presents a considerable challenge, even for experienced radiologists, due to factors like varying lesion characteristics and the presence of normal background tissue.
Current deep learning approaches for lesion segmentation have made strides, but they often fall short in real-world clinical scenarios. These models typically focus on larger lesions and overlook two vital pieces of information that radiologists routinely use: longitudinal imaging data (previous scans of the same patient) and clinical assessments, such as the BI-RADS score, which indicates the likelihood of malignancy.
Introducing LesiOnTime: A Novel Approach
To bridge this gap, researchers have developed LesiOnTime, a new 3D segmentation approach designed to mimic the diagnostic workflow of radiologists. LesiOnTime uniquely integrates both longitudinal imaging and BI-RADS scores to enhance the accuracy of small breast lesion segmentation. This innovative model aims to provide more reliable early lesion detection, especially in high-risk patients undergoing regular screening.
The core of LesiOnTime lies in two key components:
-
Temporal Prior Attention (TPA) Block: This component dynamically integrates information from a patient’s previous and current MRI scans. Unlike some prior longitudinal methods, LesiOnTime’s TPA block does not require manual pixel-wise annotations for earlier timepoints. Instead, it intelligently learns to weigh the relevance of the longitudinal scans, focusing on meaningful temporal changes while preserving the essential features of the current scan. This allows the model to adaptively emphasize subtle or slowly evolving lesions over time.
-
BI-RADS Consistency Regularization (BCR) Loss: This loss function embeds clinical domain knowledge directly into the model’s training process. It ensures that the model’s internal representations (latent space) align with clinical progression indicated by BI-RADS scores. For instance, if BI-RADS scores remain stable between visits, the BCR loss encourages feature similarity, reflecting minimal pathological change. Conversely, if scores indicate progression (e.g., a new enhancement or lesion growth), the loss allows for greater feature divergence, mirroring the clinical reality.
Also Read:
- Enhancing Medical Image Translation with Dynamic Test-Time Adaptation
- Smart AI Agents Boost Accuracy in Radiology Diagnostics
Performance and Impact
LesiOnTime was evaluated on a carefully curated in-house dataset of high-risk patients with DCE-MRI scans. The results demonstrated that LesiOnTime significantly outperforms existing state-of-the-art single-timepoint and longitudinal segmentation models. It achieved a higher Dice score (a common metric for segmentation accuracy) and superior precision, recall, and Hausdorff distance, indicating better boundary delineation and lesion localization, particularly for extremely small and challenging lesions.
Ablation studies, where components of LesiOnTime were individually removed, confirmed that both the TPA block and the BCR loss contribute complementary performance gains, highlighting their importance to the model’s success. The research also showed that the BCR effectively guides the model’s internal features to cluster according to BI-RADS scores, reinforcing its ability to learn clinically relevant patterns.
While the study was conducted at a single center, the principles behind LesiOnTime – leveraging temporal and radiological priors – suggest its potential for transferability to other clinical settings. Importantly, LesiOnTime simplifies clinical deployment by not requiring manual lesion annotations for prior timepoints and using BI-RADS scores only during the training phase, not during actual inference.
This work represents a significant step forward in automated breast lesion segmentation, offering a more reliable and context-aware approach for early breast cancer screening. For more detailed information, you can refer to the full research paper: LesiOnTime – Joint Temporal and Clinical Modeling for Small Breast Lesion Segmentation in Longitudinal DCE-MRI.
