اميرحسين سالك نيا

Since the last decade, advances in modern hardwares and the development of large-scale datasets have made deep neural networks easier to train and infer. Taking inspiration from the structure of the human brain, deep learning models achieve phenomenal performance across a wide range of artificial intelligence applications. It is now a common practice to use these models in a variety of domains such as image classification, face recognition, object detection, medical image analysis, and robot image processing. In the field of medical image analysis, powerful methods of segmenting images are important for identifying tissues and delineating essential interpretations. There are, however, obstacles to deploying deep neural networks developed for image segmentation in a real-world scenarios where hardware capabilities are limited. Because deep neural networks require powerful hardware due to their high computational load, even during the inference process, this issue prevents the development of comprehensive computer-aided detection tools that could be compatible with hardware platforms with limited computing power. Therefore in this project, we propose a method based on knowledge distillation to introduce a model with confined computational load, but with similar accuracy as contemporary models. In this research to eva‎luate the proposed method, two abdominal CT-scan datasets are utilized: Synapse and CT-1K. On these two datasets, this method is able to increase the accuracy of the light weight model, Enet, by 11 and 49 percent, respectively.

يادگيري عميق , بينايي ماشين , قطعه بندي تصاوير , پردازش تصاوير پزشكي , چكانش دانش

Deep learning , Machine vision , Image segmentation , Medical image analysis , Knowledge distillation

Amir Hossein Saleknia

Ahmad Ayatollahi