Deep Superpixel Generation and Clustering for Weakly Supervised Segmentation of Brain Tumors in MR Images

TL;DR

The paper tackles weakly supervised brain tumor segmentation in MR images by jointly learning a deep superpixel generator and a superpixel clustering model guided by localization seeds derived from a binary tumor classifier and RISE heatmaps. Operating on 2D BraTS 2020 slices, the method achieves competitive Dice scores (≈0.69–0.73) and HD$_{95}$ (≈13–18) on the test set, while delivering substantially faster inference than traditional superpixel baselines. Key contributions include the simultaneous generation and clustering of deep superpixels, the use of seed-based priors to guide segmentation without pixel-level ground truth, and demonstrated generalization to BraTS 2023 data. The approach enables efficient, initial tumor segmentations suitable for radiologist refinement or semi-supervised downstream tasks, with notable limitations in 2D scope and seed dependence that invite future 3D extension and end-to-end training.

Abstract

Training machine learning models to segment tumors and other anomalies in medical images is an important step for developing diagnostic tools but generally requires manually annotated ground truth segmentations, which necessitates significant time and resources. This work proposes the use of a superpixel generation model and a superpixel clustering model to enable weakly supervised brain tumor segmentations. The proposed method utilizes binary image-level classification labels, which are readily accessible, to significantly improve the initial region of interest segmentations generated by standard weakly supervised methods without requiring ground truth annotations. We used 2D slices of magnetic resonance brain scans from the Multimodal Brain Tumor Segmentation Challenge 2020 dataset and labels indicating the presence of tumors to train the pipeline. On the test cohort, our method achieved a mean Dice coefficient of 0.691 and a mean 95% Hausdorff distance of 18.1, outperforming existing superpixel-based weakly supervised segmentation methods.

Figures (2)

• Figure 1: Flowchart of proposed weakly supervised segmentation method. For the localization seeds component; green indicates positive seeds, magenta indicates negative seeds, black indicates unseeded regions. Solid lines represent use as inputs and outputs. Dashed lines represent use in loss functions.
• Figure 2: Visualization of T2-FLAIR channel of MR images, generated superpixels, output segmentations, and true segmentations for three examples.