Ensemble Model of U-Net EfficientNet-B3, U-Net EfficientNet B6, CoaT, SegFormer for Segmenting Functional Tissue Units in Various Human Organs
G.V.S. Akash Bharadwaj, Y. Ramya Sree, Jammigumpula Lakshmi Varshita, Srilatha Chebrolu · 2023
Accurately identifying and segmenting functional tissue units is critical for comprehending the structure and function of human organs. This work presents a pioneering method for identifying and semantically segmenting functional tissue units in diverse human organs using an ensemble model. The proposed approach initially extracts features from medical images using a pre-trained U-Net with a backbone as EfficientNet-B3, U-Net with a backbone as EfficientNet-B6, SegFormer MiT-B3, CoaT lite small, and CoaT lite medium, which is then ensembled and utilized to forecast the pixel-wise segmentation mask. Additionally, attention mechanisms were integrated to capture the contextual dependencies of tissue units within the images. To assess the performance of the proposed ensemble model, images from data sources Human Bio-Molecular Atlas Program (HuBMAP) and Human Protein Atlas (HPA) consisting of biopsy slides from various organs were considered. Experiments indicate that the proposed ensemble model outperforms individual models when evaluated using the Dice Co-efficient metric. This innovative approach has the potential to be utilized in other medical imaging tasks that involve functional tissue units for disease diagnosis and treatment planning.