State-of-the-art multi-modal brain tumor segmentation methods often rely on large quantities of manually annotated data to produce acceptable results. In settings where such labeled data may be scarce, there may be value in exploiting cheaper or more readily available data through clinical trials, such as Response Assessment in Neuro-Oncology (RANO).
This study demonstrates the utility of such measurements for multi-modal brain tumor segmentation, whereby an encoder network is first trained to regress synthetic “Pseudo-RANO” measurements using a mean squared error loss with cosine similarity penalty to promote orthogonality of the principal axes. Using oriented bounding-boxes to measure overlap with the ground truth, we show that the encoder model can reliably estimate tumor principal axes with good performance. The trained encoder was combined with a randomly initialized decoder for fine-tuning as a U-Net architecture for whole tumor (WT) segmentation.
Our results demonstrate that weakly supervised encoder models converge faster than those trained without pre-training and help minimize the annotation burden when trained to perform segmentation.
The use of cheap, low-fidelity labels in the context allows for both faster and more stable training with fewer densely segmented ground truth masks, which has potential uses outside this particular paradigm.