Investigation of ensembles of deep learning models for improved chronic kidney diseases detection in CT scan images

Abstract

• This paper investigated the likelihood of ensemble of vision transformers and convolution-based deep learning models to improve the accuracy of state-of-the-art model. The focus is on the detection of kidney stone, cyst and tumor from whole-body CT scan images. • The weighted ensembles produced a marginal improvement on the state-of-the-art. • Kidney stones were the most challenging type of CKD for the models investigated. Notwithstanding, it was clear from the results that models that employ hierarchical feature learning approach have a greater confidence on the task of CKD classification. The loss of renal function is a growing public health issue that affects up to 10 % of the global population. It is particularly endemic in low income and lower-middle income countries (LI-LMIC) where poor awareness, shortage of personnel and economic challenges complicate the multifaceted effect of chronic kidney diseases (CKD). CKD is projected to be the 5th highest cause of years of life lost by 2040, globally. Thus, continuous improvement of the achievements of the performance of intelligent systems for automated diagnosis of CKD is a key strategy to alleviating the effect of inadequate number of qualified and experienced nephrologists on the public health outcomes in LI-LMIC. This research studied ensembles of three convolution-based deep learning models and three vision transformers, namely CCT, Swin transformer (SwinT), EANet, VGG16, Inception-v3 and Resnet50, for improved detection of renal diseases - kidney stone, tumour and cyst using whole body CT scan images. Our experiments were carried out in two stages, first, the performance of the six models were assessed to establish a baseline. Thereafter, two ensemble configurations – Inception-v3-CCT-SwinT and VGG16-EANet-ResNet50 - were studied. SwinT outperformed all the models in the baseline experiments with an overall accuracy of 99.52 %. In a similar manner, the weighted ensemble involving the SwinT achieved the highest accuracy of 99.67%, representing an improvement of 0.15 %. This improvement can be considered an important achievement in the domain of medical diagnosis where it can directly result in improved diagnostic accuracy and reduced errors. In general, nonetheless, kidney stone was the most difficult disease to detect for all the models investigated. This outcome aligns with literature and suggests that further research effort is needed to address this challenge.