A Lightweight Deep Semi-supervised Student Model for Medical Image Segmentation

Medical image segmentation plays a vital role in healthcare, serving as an indispensable tool for delineating anatomical structures essential for diagnostic accuracy and treatment planning. And it is also an interesting computer vision challenge. Deep neural networks have shown significant effectiveness and extensive utilization in this task. Researchers and practitioners often fine-tune pre-trained models on their specific datasets to achieve better performance with less labeled data and computational resources. Data collection and labeling of medical data for the training process are still limited, models are becoming increasingly deep and complex, demanding more computational resources. With the widespread application of medical imaging on low-cost edge and mobile devices, there is a growing need to deploy segmentation models efficiently, especially in remote or resource-constrained areas. This necessitates research into maintaining performance when using a lightweight model with limited labeled samples. The study explores and implements an effective learning strategy aimed at constructing a polyp segmentation model that strikes a balance between complexity and performance. BKAI-IGH NeoPolyp-Small dataset is used with 1000 WLI and 200 FICE neoplastic and non-neoplastic images. The evaluation metrics used are the Dice score and the IoU score. Our experiments involve refining a baseline model with 6.6 million parameters and 31.2 million parameters reference model, which yields an improved 10% Dice score performance compared to previous iterations and achieves 99% performance comparable to the reference model.