Tumor Delineation For Brain Radiosurgery by a ConvNet and Non-Uniform Patch Generation

Egor Krivov, Valery Kostjuchenko, Alexandra Dalechina, Boris Shirokikh, Gleb Makarchuk, Alexander Denisenko, Andrey Golanov, Mikhail Belyaev

Abstract

Stereotactic radiosurgery (SRS) demands highly precise tumor delineation to maximize therapeutic efficacy while sparing healthy brain tissue. This paper adapts the DeepMedic 3D convolutional neural network (CNN) to address the specific challenges of segmenting small brain tumors. We introduce two novel non-uniform patch-sampling strategies: Tumor Sampling (TS), which balances the representation of small and large lesions, and High Intensity Sampling (HIS), which reduces false positives from bright anatomical structures. Our results demonstrate significant improvements in sensitivity for small metastases and a reduction in false positives for meningiomas, while maintaining temporal robustness across different imaging eras.

Introduction

The clinical application of deep learning in SRS is often hindered by two primary issues. First, the "Small Lesion Problem": standard uniform voxel sampling causes models to be biased toward larger tumors, leading to high false-negative rates for small metastases. Second, the "False Positive Problem": bright non-tumor structures, such as the skull or specific membranes, often mimic the appearance of contrast-enhanced tumors in MRI. This study proposes a modified training regime for the DeepMedic architecture to overcome these hurdles without requiring computationally expensive preprocessing steps like skull stripping.

Methods

We utilized the DeepMedic architecture, an 11-layer 3D CNN featuring parallel pathways for local high-resolution and global low-resolution context. To improve performance, we modified the training process with two sampling techniques:

  • Tumor Sampling (TS): Instead of uniform voxel selection, the algorithm first selects a tumor instance and then samples a voxel within it. This ensures that small lesions (e.g., 5mm) are represented as frequently as large ones (e.g., 3cm) during training.
  • High Intensity Sampling (HIS): This form of hard negative mining targets background voxels with intensities above the 90th percentile. By forcing the model to train on these "difficult" bright regions with a 30% probability, the model learns to distinguish tumors from bright anatomical structures like the skull.

The model was trained on a historic dataset (2005–2011) comprising 404 metastasis, 341 meningioma, and 252 schwannoma patients. Evaluation was performed using 5-fold cross-validation and a separate "modern" hold-out set from 2017 to test temporal stability. All experiments used T1-weighted contrast-enhanced (T1c) MRI scans.

Key Results

The integration of TS and HIS yielded substantial performance gains. For metastases, sensitivity improved from 0.898 to 0.946 in cross-validation. On the 2017 hold-out set, TS increased sensitivity from 0.856 to 0.912, confirming the model's ability to generalize across different imaging hardware and time periods. For meningiomas, the HIS method effectively reduced the average number of false positives per image from 6.9 to 4.6.

Segmentation results comparing a large tumor (left) and a small metastasis (right) across T1c, Ground Truth, and Prediction.

A key insight of this work is that class imbalance in medical imaging is often an instance-level problem rather than just a voxel-level one. By shifting to instance-based sampling, the model becomes significantly more sensitive to life-threatening small lesions. Furthermore, by avoiding "brain extraction" preprocessing, the proposed workflow remains efficient for real-time clinical use.

Discussion and Conclusion

Our findings suggest that non-uniform sampling strategies are essential for training robust segmentation models in radiosurgery. While the current approach relies on a single MRI modality (T1c), it achieves high sensitivity and temporal robustness. Future work will investigate the inclusion of additional sequences like T2 or FLAIR to further suppress false positives. Additionally, the trade-off between sensitivity and false positive rates remains a key area for optimization, potentially through the use of separate detection and segmentation stages in the clinical pipeline.

Summary updated

BibTeX
@inproceedings{krivov2018tumor,
  author = {Egor Krivov and Valery Kostjuchenko and Alexandra Dalechina and Boris Shirokikh and Gleb Makarchuk and Alexander Denisenko and Andrey Golanov and Mikhail Belyaev},
  title = {{Tumor Delineation For Brain Radiosurgery by a ConvNet and Non-Uniform Patch Generation}},
  booktitle = {Patch-MI @ MICCAI},
  year = {2018},
  url = {https://openreview.net/forum?id=WBvuPsqnoeh},
}
Figure