Kidney Abnormalities Prediction in Ultrasound Images using Transfer Learning Approach

Transfer Learning Algorithm is becoming popular for medical image analysis, because it has high level of flexibility in dealing with Image dataset with minimum number of Images. It makes convenient for researchers to test and train medical images, assign the determined weights thus reducing the computational complexities and pre-processing time. Presence of stones, tumour, cyst and other kidney abnormalities in Kidney Ultrasound Images is a challenging task for a radiologist. There exists a difference of opinion between an experienced physician and a beginner in the field of urology and nephrology. Requirement of an intelligent tool for diagnosis in the medical field has increased in the recent years due to the enormous growth of Artificial Intelligent and Machine Learning algorithms. In this proposed work Pretrained Efficient Net models B1-B5 are used with the pretrained weights of ImageNet data set thus, reducing the pre-processing time taken for random initialization of weights. Ultrasound Kidney Image Dataset are taken as inputs for training phase and testing Phase. Fine Tuning is adapted by changing the hyper parameter setting. Performance metrics such asF1Score and AUC are evaluated. Efficient Net B4 Model gives satisfactory result in terms of AUC as 89.95% and F1 Score 82.9%. Cross validation when carried out by training the models with predetermined weights of CT scan Kidney Images resulted in improving AUC to 90.9% for the prediction of kidney stones, cyst and tumour.