The Challenge
ML for finding patterns in brain cells studies
Brain cancer (and related conditions) can be described very generally as the result of cells in the brain losing normal “growth control” and starting to divide or behave in uncontrolled ways. That loss of control usually comes from changes in how the cell’s DNA is regulated—either damage/mutations in genes or misregulation of gene activity (which genes are turned on/off, and when). Solvd’s RnD team was tasked with a project to help a team of scientists to analyze small samples of brain cells, and use machine learning to find subtle patterns in complex biological imaging/data that are hard to see by eye, so they can spot early signs of disorganization or abnormal behavior and better understand the pathways that may lead to tumor formation. Small samples create a constant risk of overfitting and makes every single data decision disproportionately consequential. Images from different experiments varied significantly in brightness, contrast, and imaging artifacts, which meant heavy preprocessing had to happen before any training could begin at all. Two different imaging modalities, captured separately, needed precise spatial alignment before they could be used together, and any imperfection in that alignment introduced noise that was especially damaging given how small the structures being studied actually were. And because most modern AI vision models are built for 2D images, extending them to 3D volumetric microscopy data required real engineering work across model architecture, memory management, and spatial encoding, not just a change of input format.
