TLOD: Innovative ovarian tumor detection for accurate multiclass classification and clinical application

Abstract

Ovarian tumors pose a major threat to women's health, mostly remaining undetected until they reach advanced stages, resulting in complex treatment and decreased survival rates. Besides, tumor heterogeneity is more responsible for poor treatment response and adverse prognosis. The purpose of this research is to identify ovarian epithelial tumors in premature stage using histopathological images. In this research, we address the need for an improved ovarian tumor detection method through the development of an innovative simple intelligent approach ‘Transfer Learning with ResNet-based Deep Learning for Ovarian Tumor Detection (TLOD)’. The innovation lies in modifying the final classification layers of the standard residual network with dense blocks, in which it merges and passes the extracted features to multi-classification of ovarian tumors, empowering medical professionals to make accurate diagnoses and improve patient outcomes. The histopathological images of ovarian epithelial tissue samples are taken from the National Cancer Institute Genomic Data Commons Data Portal for analysis. The variations in input data are handled by preprocessing pipelines such as resizing and intensity standardization. To increase model robustness and generalization ability, the size of training dataset is expanded using data augmentation. These enhanced images are subsequently trained by the proposed TLOD technique to detect and classify different subtypes of ovarian epithelial tumors. The performance manipulating control parameters such as batch size, learning rate, and momentum are optimized using a Stochastic Gradient Descent (SGD) optimizer. Overall multi-class ovarian tumor detection performance of 98.82% is achieved by the proposed TLOD, offering highly reliable and efficient tool for early ovarian tumor detection.