Deep Learning Based Tympanic Membrane Segmentation Using Residual Double Attention UNet

Abstract

The airy spaces of the middle ear and temporal bone are home to the Eustachian tube, which is covered by a mucous membrane that is infected and inflamed in Otitis Media (OM). Another of the most prevalent diseases is OM. Otoscope pictures are visually inspected in clinical settings to make the diagnosis of OM. Being subjective and prone to mistakes, this procedure is susceptible. In this study a unique framework Hybrid Colour Residual Double Attention UNet (HCRDAUNet)  model is proposed to effectively segment the Tympanic membrane. This model utilizes the strength of three different colour spaces namely RGB, LUV and HSV into a single joined semantic segmentation model with attention mechanism. The proposed attention gate in this approach applies the gating outcome on two different scales of feature map to accurately localize the eardrum. The proposed HCRDAUNet model archives up to 96% of dice co-efficient and 95% of F1-score, which shows that the proposed model attains significant improvements in performance, compared to state art of semantic segmentation models.