This website requires JavaScript.

Knowledge Distillation for Adaptive MRI Prostate Segmentation Based on Limit-Trained Multi-Teacher Models

Eddardaa Ben LoussaiefHatem RashwanMohammed AyadMohammed Zakaria HassanDomenec Puig
Mar 2023
With numerous medical tasks, the performance of deep models has recentlyexperienced considerable improvements. These models are often adept learners.Yet, their intricate architectural design and high computational complexitymake deploying them in clinical settings challenging, particularly with deviceswith limited resources. To deal with this issue, Knowledge Distillation (KD)has been proposed as a compression method and an acceleration technology. KD isan efficient learning strategy that can transfer knowledge from a burdensomemodel (i.e., teacher model) to a lightweight model (i.e., student model). Hencewe can obtain a compact model with low parameters with preserving the teacher'sperformance. Therefore, we develop a KD-based deep model for prostate MRIsegmentation in this work by combining features-based distillation withKullback-Leibler divergence, Lovasz, and Dice losses. We further demonstrateits effectiveness by applying two compression procedures: 1) distillingknowledge to a student model from a single well-trained teacher, and 2) sincemost of the medical applications have a small dataset, we train multipleteachers that each one trained with a small set of images to learn an adaptivestudent model as close to the teachers as possible considering the desiredaccuracy and fast inference time. Extensive experiments were conducted on apublic multi-site prostate tumor dataset, showing that the proposed adaptationKD strategy improves the dice similarity score by 9%, outperforming all testedwell-established baseline models.