Automated Detection of Cystitis in Ultrasound Images Using Deep Learning Techniques

The proposed method aims to estimate the urinary Bladder Wall Thickness (BWT) from ultrasound (US) images to detect cystitis. Our method proposes a novel deep learning algorithm that segments the Bladder Wall from the ultrasound images of the urinary bladder, following which feature extraction and classification are performed to categorize the images as presence or absence of cystitis. The proposed study focused on a CYSNET CNN (Convolutional Neural Network) model for detecting cystitis in the urinary bladder and compares its accuracy with a transfer learning-based pre-trained model like ResNet50 and state-of-the-art Vision Transformer. Among the total population studied (N=250), 125 subjects with cystitis and 125 normal subjects, were included. The bladder wall thickness of cystitis was segmented using the U-Net semantic segmentation model. Eight features constituting contour and thickness were extracted from the segmented bladder wall. The best five features were selected using the Univariate feature selection method based on ANOVA F statistics as the scoring scale. The selected five features were classified into cystitis and normal using three different Machine Learning (ML) Classifiers such as AdaBoost, RepTree, and NaÏve Bayes. Three different CYSNET models with varying convolution layers were developed to detect cystitis in ultrasound images. The performance of the CYSNET models is compared with the ML classifiers, ResNet 50 model, and Vision Transformer. The CYSNET model 3 outperformed with the classification accuracy of 95% compared to the Adaboost network (90%), ResNet50 model (88.7%) and Vision Transformer (92.1%). Hence, the developed CYSNET model could be used as a computer-aided diagnostic tool for the detection of cystitis in ultrasound images.