Clinical Inspired MRI Lesion Segmentation

TL;DR

The paper tackles MRI lesion segmentation across pre- and post-contrast T1-weighted sequences, a challenging task due to diverse enhancement patterns. It introduces a residual fusion framework with a dual-branch encoder (main post-contrast and auxiliary pre-contrast) and a multi-scale fusion strategy, formalized by the residual block $H(x)=E(x_{ ext{main}})+x_{ ext{aux}}$. A lightweight dynamic weighting mechanism using a 1×1×1 convolution optimizes auxiliary contributions without substantial parameter overhead, enabling robust performance across scales. Empirically, the method achieves state-of-the-art results on BraTS2023 brain tumor segmentation and an in-house breast MRI dataset, with clear improvements in small or irregular lesions and reduced false positives, underscoring its clinically relevant potential.

Abstract

Magnetic resonance imaging (MRI) is a potent diagnostic tool for detecting pathological tissues in various diseases. Different MRI sequences have different contrast mechanisms and sensitivities for different types of lesions, which pose challenges to accurate and consistent lesion segmentation. In clinical practice, radiologists commonly use the sub-sequence feature, i.e. the difference between post contrast-enhanced T1-weighted (post) and pre-contrast-enhanced (pre) sequences, to locate lesions. Inspired by this, we propose a residual fusion method to learn subsequence representation for MRI lesion segmentation. Specifically, we iteratively and adaptively fuse features from pre- and post-contrast sequences at multiple resolutions, using dynamic weights to achieve optimal fusion and address diverse lesion enhancement patterns. Our method achieves state-of-the-art performances on BraTS2023 dataset for brain tumor segmentation and our in-house breast MRI dataset for breast lesion segmentation. Our method is clinically inspired and has the potential to facilitate lesion segmentation in various applications.

Figures (5)

• Figure 1: Overview of enhancement patterns. The four columns represent pre-contrast images, post-contrast images, the lesions drawn by radiologists, and the subtraction images (SUB). The three rows represent fibroadenoma, typical mass, and well-vascularized glandular tissue respectively.
• Figure 2: Left: The architecture of our network, featuring a main branch and an auxiliary branch, each comprising encoding blocks, as well as fusion weight learning blocks. Right: residual modules of different versions. (a) is the original version from ResNet, (b) is our modified version which incorporates two branches, and (c) is our final residual module which adds a fusion weight learning block into (b).
• Figure 3: Visualization of breast lesion segmentation. Each row represents a case from our in-house dataset.
• Figure 4: Visualization of atypical enhanced lesion.
• Figure 5: Visualization of brain tumor segmentation. Each row shows a case from the BraTS2023 dataset. WT=red+green+blue; TC=red+green; EN=red.