Addressing data imbalance challenges in oral cavity histopathological whole slide images with advanced deep learning techniques

Abstract

Oral Cavity Squamous Cell Carcinoma (OCSCC) represents a common form of head and neck cancer originating from the mucosal lining of the oral cavity, often detected in advanced stages. Traditional detection methods rely on analyzing hematoxylin and eosin (H&E)-stained histopathological whole-slide images, which are time-consuming and require expert pathology skills. Hence, automated analysis is urgently needed to expedite diagnosis and improve patient outcomes. Deep learning, through automated feature extraction, offers a promising avenue for capturing high-level abstract features with greater accuracy than traditional methods. However, the imbalance in class distribution within datasets significantly affects the performance of deep learning models during training, necessitating specialized approaches. To address the issue, various methods have been proposed at both data and algorithmic levels. This study investigates strategies to mitigate class imbalance by employing a publicly available OCSCC imbalance dataset. We evaluated undersampling methods (Near Miss, Edited Nearest Neighbors) and oversampling techniques (SMOTE, Deep SMOTE, ADASYN) integrated with transfer learning across different imbalance ratios (0.1, 0.15, 0.20, 0.30). Our findings demonstrate the effectiveness of SMOTE in improving test performance, highlighting the efficacy of strategic oversampling combined with transfer learning in classifying imbalanced medical datasets. This enhances OCSCC diagnostic accuracy, streamlines clinical decisions, and reduces reliance on costly histopathological tests.