A new AI-powered medical imaging technique can predict the precise shape of bone removal required during cochlear implant surgery using only pre-operative CT scans. Developed by researchers from St. Mary's University, Trinity University, Vanderbilt University, and the Center for Advanced AI, the method leverages two advanced machine learning approaches to learn from "before and after" scans without relying on time-consuming manual labels. Cochlear implants are life-changing devices for people with severe hearing loss, but their placement requires a delicate surgical step called mastoidectomy—removing part of the bone behind the ear to access the inner ear. The resulting cavity varies significantly between patients and lacks a clear external boundary, making it challenging for traditional imaging tools to predict accurately. Reliable pre-surgical predictions could enhance surgical planning, support robotic assistance, improve navigation systems, and ultimately lead to better patient outcomes. For years, researchers have struggled to develop AI tools capable of consistently predicting the mastoidectomy shape. The new method overcomes this challenge by using a two-stage learning process that works effectively even when high-quality labeled data is unavailable. The first stage compares pre-surgery CT scans with post-surgery scans. Despite the noise and variability in post-surgery images, the AI uses a structure-focused mathematical comparison to identify patterns in bone removal, learning automatically without expert annotations. The predictions from this initial model are then used as "weak labels" to train a second, more refined model. This second model employs a specialized 3D loss function based on the Student-t distribution, which is robust to outliers and imperfect data. This step significantly improves the accuracy and reliability of the final bone-removal prediction. The system was tested on 751 pairs of pre- and post-surgery CT scans. When evaluated against 32 manually labeled examples from expert surgeons, the AI achieved a mean Dice score of 0.72—a strong result that surpasses several widely used medical imaging models. A higher Dice score indicates a closer match between the predicted and actual bone cavity shapes. The researchers also demonstrated the ability to generate a 3D model of the predicted post-surgery bone surface. Such models could one day assist surgeons in real time during operations or serve as training tools for medical students. This breakthrough is significant because it offers a practical way to train AI systems in medical imaging when detailed manual labeling is impractical or too costly. Many anatomical structures are too complex to label by hand, and this approach could extend to other surgical procedures with similarly intricate shapes. For patients, the technology holds promise for safer, more efficient cochlear implant surgeries. It could enable better pre-operative planning, support the integration of robotic systems, and improve intraoperative guidance. While the results are encouraging, the team cautions that further validation across diverse clinical settings is needed before the tool can be adopted in routine care. They also aim to enhance the 3D models with more realistic textures to improve usability during live procedures.