A project by students from Oaklands School, in collaboration with Orbyts Fellow Clea Dronne from UCL Medical Physics & Biomedical Engineering, has developed an artificial intelligence (AI) model to segment organs-at-risk (OARs) for radiotherapy planning. Radiotherapy, which uses high-energy rays to target cancer cells, carries the risk of harming nearby healthy organs. The current planning process, which involves identifying and segmenting tumors and OARs from medical images, is often complex, slow, and time-consuming. This new AI model aims to automate this crucial step, thereby improving the accuracy and safety of cancer treatment.

The research utilised a U-Net convolutional neural network, a deep learning model specifically designed for image segmentation. The training process involved feeding 3D images into the U-Net, which then processed them through various layers to extract and combine features, ultimately producing an output image where different regions (e.g., organs or tumours) were labeled with distinct classes. The model learned by comparing its predicted segmentations to the real, "ground truth" segmentations and iteratively reducing its error, or "loss," over multiple epochs using batches of data.

The results demonstrated the model's significant learning capability. To evaluate the model's performance, the Dice Score was employed, a metric that measures the similarity between the predicted and ground truth segmentations. The Dice scores were very close to 1 for most classes, signifying high accuracy in segmentation. This successful development of a deep learning model for segmenting OARs in CT scans on synthetic data holds considerable promise for enhancing real-world radiotherapy planning, making it quicker, more accurate, and ultimately safer for patients!