Deep Machine Learning Histopathological Image Analysis for Renal Cancer Detection
Jia Chun Koo, Yan Chai Hum, Khin Wee Lai, Wun‐She Yap, Swaminathan Manickam, Yee Kai Tee · 2022
Renal cancer is one of the top causes of cancer-related deaths among men globally. Early detection of renal cancer is crucial because it can significantly improve the probability of survival rate. However, assessing the histopathological renal tissues is a labor-intensive job and traditionally, this is done manually by a pathologist, leading to a high possibility of misdetection and/or misdiagnosis especially in the early stages and prone to inter-pathologist variations. The development of an automatic histopathological diagnosis of renal cancer can greatly reduce the bias and provide accurate characterization of diseases even though the nature of pathology and microscopy are highly complex and complicated. This paper investigated the use of deep learning methods to develop a binary histopathological image classification model (cancer or normal). 783 whole slide images of renal tissue were processed into patches using PyHIST tool at 5x magnification power before feeding them to the deep learning models. Five pre-trained deep learning architectures, namely VGG, ResNet, DenseNet, MobileNet, and EfficientNet, were trained with transfer learning on the CPTAC-CCRCC dataset and their performances were evaluated. EfficientNetB0 achieved the state-of-the-art accuracy (97%), specificity (94%), F1-score (98%) and AUC (96%) but slightly inferior recall (98%) when compared to the best published results in the literature. These findings showed that the proposed deep learning approach can effectively classify the histopathological images of renal tissue into tumor and non-tumor classes to make pathology diagnosis more efficient and less labor intensive.