Deep learning for histopathological image cancer detection

Gan, Zi Hui (2025) Deep learning for histopathological image cancer detection. Final Year Project, UTAR.

PDF Download (2195Kb)

Abstract

Head and neck cancers (HNC) are among the most prevalent cancers globally, with high mortality and poor prognosis often resulting from late-stage diagnoses. However, diagnostic difficulties are compounded by the histological complexity of HNCs and the subjective nature of manual histopathological analysis, which is prone to human error and inter-observer variability. Therefore, this study proposed a deep learning approach to assist in the classification of HNC from histopathological whole slide images, aiming to improve diagnostic accuracy and reduce observer bias. This study adopted the Head and Neck Squamous Cell Carcinoma dataset from the Clinical Proteomic Tumor Analysis Consortium, which consists of 390 whole slide images from various head and neck cancer sites, including 122 benign and 268 tumor slides. Convolutional neural network (CNN) models were trained using a transfer learning strategy, incorporating variants from the DenseNet, EfficientNet, MobileNet, ResNet, and VGG families. These models were fine-tuned using pre-trained weights and further evaluated for classification performance at three magnification levels (1.25×, 2.5×, and 5×). The top-performing CNN models were then combined using ensemble learning techniques to improve overall accuracy and robustness. The ensemble approach, particularly the majority voting with five models ensemble, outperformed individual models, achieving an accuracy of 96.09%, along with improved performance in sensitivity, precision, and F1-score. Visual interpretability tools, such as Gradient-weighted Class Activation Mapping, were employed to provide insights into the models' decision-making processes, enhancing the transparency and trustworthiness of the artificial intelligence predictions. The study also compared the CNN-based models to Vision Transformer models, showing that CNN ensembles achieved superior performance in classification tasks. This research highlights the potential of deep learning, particularly ensemble methods, in histopathological image analysis, with significant applications in computer-aided diagnosis for cancer detection. Further work should focus on addressing class imbalance, integrating the models into a clinical pipeline, and exploring multimodal learning to enhance model performance and clinical applicability. Keywords: deep learning, convolutional neural networks, ensemble learning, whole slide image, head and neck cancer, Subject Area: Q300-390 Cybernetics

Actions (login required)