Anatomy-Informed Deep Learning and Radiomics for Automated Neurofibroma Segmentation in Whole-Body MRI

TL;DR

The study tackles automated neurofibroma segmentation in fat-suppressed T2-weighted WB-MRI for NF1 by introducing a three-stage pipeline that couples anatomy-aware context with radiomics-based tumor candidate classification. Stage 1 refines anatomy masks and adds an NF high-risk zone, Stage 2 uses an ensemble of 3D anisotropic anatomy-informed U-Nets to generate a segmentation confidence mask, and Stage 3 extracts radiomic features from tumor candidates and classifies them with region-specific random forests. The approach yields substantial gains in segmentation and detection, notably a 68% per-scan DSC improvement and a 21% per-tumor DSC increase in high-burden cases, along with up to a two-fold F1 gain for tumor detection, while revealing sensitivity to domain shifts and low-tumor-burden scenarios. The pipeline is integrated into 3D Slicer with open-source code, offering a practical route toward clinical deployment, and the authors discuss limitations of current evaluation protocols and the need for harmonization across MRI protocols.

Abstract

Neurofibromatosis Type 1 is a genetic disorder characterized by the development of neurofibromas (NFs), which exhibit significant variability in size, morphology, and anatomical location. Accurate and automated segmentation of these tumors in whole-body magnetic resonance imaging (WB-MRI) is crucial to assess tumor burden and monitor disease progression. In this study, we present and analyze a fully automated pipeline for NF segmentation in fat-suppressed T2-weighted WB-MRI, consisting of three stages: anatomy segmentation, NF segmentation, and tumor candidate classification. In the first stage, we use the MRSegmentator model to generate an anatomy segmentation mask, extended with a high-risk zone for NFs. This mask is concatenated with the input image as anatomical context information for NF segmentation. The second stage employs an ensemble of 3D anisotropic anatomy-informed U-Nets to produce an NF segmentation confidence mask. In the final stage, tumor candidates are extracted from the confidence mask and classified based on radiomic features, distinguishing tumors from non-tumor regions and reducing false positives. We evaluate the proposed pipeline on three test sets representing different conditions: in-domain data (test set 1), varying imaging protocols and field strength (test set 2), and low tumor burden cases (test set 3). Experimental results show a 68% improvement in per-scan Dice Similarity Coefficient (DSC), a 21% increase in per-tumor DSC, and a two-fold improvement in F1 score for tumor detection in high tumor burden cases by integrating anatomy information. The method is integrated into the 3D Slicer platform for practical clinical use, with the code publicly accessible.

Figures (7)

• Figure 1: Fat-suppressed T2-weighted whole-body MRI of two patients with varying tumor burden. (a) 34-year-old male with a tumor burden of 1,591 cm$^{3}$, showing plexiform neurofibromas (PNFs, filled arrowheads) and subcutaneous neurofibromas (empty arrowheads). (b) 22-year-old female with a tumor burden of 96 cm$^{3}$, showing PNFs (arrowhead).
• Figure 2: Proposed pipeline for automated neurofibroma (NF) segmentation, consisting of three main stages: (1) The input T2-weighted whole-body MRI (T2w WB-MRI) is segmented using the MRSegmentator model, followed by post-processing to refine the anatomy segmentation mask. (2) The input T2w WB-MRI and the refined anatomy segmentation mask are processed by an ensemble of 3D anisotropic anatomy-informed U-Net models to produce an NF segmentation confidence mask. (3) The input T2w WB-MRI, the anatomy segmentation mask, and the NF confidence mask are used to extract tumor candidates with their radiomic features. The tumor candidates are grouped by anatomical regions (head and neck, chest, abdomen, legs) and classified using random forest classifiers. The final NF segmentation mask is formed by filtering the tumor candidates.
• Figure 3: Comparison of initial and refined anatomy segmentation masks: (a) Initial coarse anatomy segmentation mask generated by MRSegmentator. (b) Refined mask with redundant anatomies removed, labels merged, neurofibroma (NF) high-risk zone added (red structure around the lungs and spine), and anatomical regions (red arrows) defined by anatomical landmarks: the highest and the lowest points of the lungs, and the lowest point of the hips (red dots). (c) T2-weighted whole-body MRI with updated anatomy segmentation masks and NF ground truth (green), highlighting NFs within the NF high-risk zone (red rectangle).
• Figure 4: Distribution of tumors across anatomical regions in four datasets: (a) Train set; (b) Test set #1; (c) Test set #2; (d) Test set #3. Each dataset is represented by a stylized human figure divided into anatomical regions. The numerical values within each region indicate the average absolute number of tumors in that region. The color of each region represents the proportion of tumors in that region relative to the total tumor count in each dataset, as indicated by the color bar.
• Figure 5: Mean F1 scores for tumor detection across anatomical regions for three test sets #1, #2, #3 (rows), and four methods (columns): (a) Method 1 (3D isotropic U-Net), (b) Method 2 (3D anisotropic U-Net), (c) Method 3 (ensemble of 3D anisotropic anatomy-informed U-Nets), (d) Method 4 (Method 3 with tumor candidate classification).