Glossary
List of terms that comes up frequently in my work.
- API: A set of routines, protocols, and tools for building software applications.
- Activation function: A mathematical function applied to each neurons output in a neural network layer to introduce non-linearity.
- BEV: Beam Eye View, a visualization perspective simulating how radiation beams 'see' the patient, aiding in plan optimization.
- Backpropagation: An algorithm used to compute the gradient of the loss function with respect to model parameters, enabling learning through gradient descent.
- Batch size: The number of training samples processed before the model's internal parameters are updated.
- Brachytherapy: Internal radiation therapy using radioactive sources placed inside or near the tumor.
- CBCT: Cone Beam CT, a 3D imaging modality integrated with LINACs to verify patient position before radiation delivery. These images are overlaid on the planning images to assess the need for repositioning.
- CI/CD: Continuous Integration / Continuous Deployment, a method for automating testing and deployment.
- CNN: Convolutional Neural Network, a type of deep neural network particularly effective for processing grid-like data such as images.
- CT Simulation: A planning CT scan used to map the patient's anatomy in treatment position for precise radiation therapy delivery.
- CTV: Clinical Target Volume, the GTV plus regions suspected of microscopic disease that need irradiation.
- Chart review: a systematic review of a patient's treatment records to ensure accuracy, safety, and compliance throughout the radiation therapy process. It is a critical quality assurance (QA) step that involves checking and verifying key elements of the treatment workflow.
- Checkpoint: Saving the models state (weights, optimizer, etc.) at regular intervals during training to allow resuming or fine-tuning later.
- Contouring: The process of delineating anatomical structures and target areas on medical images to guide radiation planning.
- Cross-entropy loss: A common loss function used for classification tasks, measuring the difference between the predicted and actual distributions.
- Cyber knife: LINAC mounted on a robotic arm that can move during treatment to track moving targets.
- DVH: Dose Volume Histogram, a graphical representation showing the distribution of radiation dose within the target and surrounding organs.
- Data augmentation: A technique used to increase the diversity of training data by applying random transformations like rotation, flipping, or scaling.
- Data split: The division of data into subsets—typically training, validation, and testing—to evaluate and generalize model performance.
- DevOps: A set of practices that combines software development and IT operations.
- Dropout: A regularization technique where randomly selected neurons are ignored during training to prevent overfitting.
- EBRT: External Beam Radiation Therapy, radiation therapy delivered from a source outside the patient's body, where the machine that never touches your body sends radiation to destroy tumors.
- ECOG: Eastern Cooperative Oncology Group Performance Status, a clinician-assigned scale from 0 to 5 that describes a patient's level of functioning in terms of self-care, daily activity, and physical ability, with higher scores indicating greater disability. It is very similar to the Karnofsky Performance Scale.
- Epoch: One complete pass through the entire training dataset by the model.
- Fine-tuning: The process of taking a pre-trained model and training it further on a new, often smaller, dataset.
- Fractionation: Dividing the total radiation dose into multiple smaller doses delivered over several sessions.
- Fully connected layer: also called a linear layer, a layer where every neuron is connected to all neurons in the previous layer, typically used in the final layers of a CNN.
- GTV: Gross Tumor Volume, visible or palpable extent of the tumor identified on imaging or clinical exam.
- Gamma knife: radioactive source producing dozens of small beams converging to a single location to treat tumours in the head/brain.
- Generalization: The model's ability to perform well on unseen data, not just the training data.
- Gradient descent: An optimization algorithm that updates model weights iteratively to minimize the loss function.
- Hypofractionation: Fewer, larger doses of radiation per session—used in some curative or palliative treatments.
- IGRT: Image-Guided Radiation Therapy, use of imaging during treatment to improve accuracy and precision of radiation delivery.
- IMRT: Intensity-modulated radiation therapy, or IMRT, (also called static IMRT) a technique where radiation beam intensity is varied and delivered from distinct angles to conform to the tumor shape and spare normal tissues. It allows people with cancer to receive higher, more effective doses of radiation while limiting damage to the healthy tissues and organs around it.
- Inference: The process of making predictions using a trained model.
- IoU: Intersection over Union, a metric used to evaluate the accuracy of object detection or segmentation by comparing overlap between prediction and ground truth.
- Isodose line: Contours that represent areas receiving the same radiation dose, used in treatment plan visualization.
- KPS: Karnofsky Performance Scale, a standardized way of measuring a cancer patient's functional status, particularly in terms of their ability to carry out daily activities. It is commonly used in radiation oncology, medical oncology, and palliative care to assess a patients general well-being and to help guide treatment decisions, prognosis, and eligibility for clinical trials. It ranges from 100 (normal), 50-60 (unable to work), 10-20 (bedridden), and 0 (dead).
- Keras: A high-level neural networks API, running on top of TensorFlow, designed for fast experimentation.
- LINAC: Linear Accelerator, a machine that generates high-energy x-rays or electrons for radiation treatment.
- Learning rate: A hyperparameter that controls how much the model weights are updated during training.
- Loss function: A function that measures the error between the predicted output and the actual output.
- MAP: Mean Average Precision, a performance metric used in object detection tasks that averages precision across recall levels.
- MLC: Multileaf Collimator, a device within the LINAC that shapes the radiation beam to conform to tumor contours.
- OAR: Organs at Risk, normal tissues sensitive to radiation, which must be protected during treatment planning.
- OIS: Oncology Information System, a digital system for managing radiation therapy workflows, documentation, and patient records.
- Overfitting: A condition where a model learns the training data too well, including noise, resulting in poor generalization.
- PTV: Planning Target Volume, CTV plus a margin to account for motion and setup variability, ensuring effective dose delivery.
- Parameters: Variables that the model learns during training, which define the model's behavior.
- Pooling layer: A layer that reduces spatial dimensions (width and height) in CNNs, typically using max or average pooling.
- Port film: x-ray images of the patient taken by the treatment beam from the perspective of the beam, in BEV (beameye view).
- Pretrained model: A model that has been previously trained on a large dataset and can be reused or fine-tuned for a different task.
- PyTorch: An open-source deep learning framework developed by Facebook, known for flexibility and dynamic computation graphs.
- QA: Quality Assurance, routine procedures to ensure safety and accuracy in radiation therapy delivery and planning.
- ReLU: Rectified Linear Unit, a widely used activation function defined as f(x) = max(0, x).
- ResNet: A deep residual network architecture known for allowing very deep networks by introducing skip connections.
- SBRT: Stereotactic Body Radiotherapy, similar to SRS, but delivered in a few (2 to 5) treatments.
- SRS: Stereotactic Radiosurgery, using highly conformal, and high dose radiation treatment usually delivered to a small target like brain metastasis in a single session (one fraction). This relies a lot on imaging, as the target is localized live, before turning on the beam.
- Segmentation: The task of partitioning an image into segments or classes, often used in medical and satellite imagery.
- Setup Error: Variability between planned and actual patient positioning, managed by imaging and margins.
- Simulation: The preparatory step involving imaging and positioning to reproduce the treatment setup consistently.
- Softmax: An activation function used in the output layer of a classification model to convert logits into probabilities.
- TPS: Treatment Planning System, software used by medical physicists and dosimetrists to design optimal radiation treatment plans.
- Tensor: A multi-dimensional array used as the basic data structure in deep learning frameworks.
- TensorFlow: An open-source deep learning framework developed by Google, known for scalability and deployment support.
- Tomotherapy: LINAC housed within a CT scanner allowing dynamic treatments based on location/size.
- Training: The process where a model learns patterns from data by updating weights to minimize the loss function.
- Transfer learning: A technique where a model developed for one task is reused as the starting point for a model on a second task.
- U-Net: A CNN architecture specifically designed for image segmentation, especially in biomedical applications.
- Underfitting: A condition where a model is too simple to capture the underlying pattern in the data.
- VMAT: Volumetric modulated arc therapy, or VMAT, (also called arc therapy) an advanced form of IMRT delivering radiation in arcs around the patient with dynamic modulation. It is a type of EBRT, and an advance over IMRT. VMAT is a newer form of EBRT that was introduced in 2007. The machine rotates around your body while youu are lying down, delivering continuous doses of radiation toward a tumor. No radiation source is placed inside your body, so there is no concern of being radioactive during or after the radiation treatments. With VMAT, you receive customized doses of radiation as a machine encircles your body in one or more rotations, or arcs.
- Validation set: A subset of data used to tune model hyperparameters and monitor performance during training.
- Weights: Parameters within the neural network that are updated during training to minimize loss.
- YOLO: You Only Look Once, a real-time object detection system that predicts bounding boxes and class probabilities directly from full images in one evaluation.