All ETDs from UAB

Advisory Committee Chair

Emily A Caffrey

Advisory Committee Members

Carlos E Cardenas

Joseph Harms

Charles A Wilson

Document Type


Date of Award



While radiotherapy has become one of the primary standards of treatment for cancer in areas such as the brain, treatment plans are limited by the time it takes physicians to contour CT images used for dose distribution calculations. Automation of the contouring process can not only improve the rate of plan formation but establish an improved standard of care. With the development and implementation of knowledge-based planning (KBP) in radiation oncology, information derived from treatment plans provide machine learning models with years of experience to assess and create future treatment plans. Since KBP model quality rely on contouring from training plans, it is crucial to assess the impact contour quality has on the model-generated plan quality. Understanding contour impact on model development can lead to more consistent, higher quality, and more optimal outcomes for the future. The brain and lung were selected as the target disease sites and 125 and 75 manual physician treatment plans were selected respectively. Scans of the disease sites then underwent autocontouring through a deep learning clinical model. Protective margins were then implemented for both the brainstem and spinal cord. Both the manual and deep learning contours were then added to the RapidPlan tool for model training. It is hypothesized that utilization of autocontouring will result in a more accurate model that allows for improved dose distribution and sparing with a higher consistency between cases than manual counterparts. Both the manual and autocontour models will be compared based on quality and consistency. While KBP and deep learning show promising potential and are already implemented in clinical settings, evaluation of treatment model training and refinement is required to further refine plan quality. Given the success of RapidPlan model plans, further development of the model would include seeking to reduce the number of manual cases needed to train KBP models.