Convolutional Neural Network for Histopathological Analysis of Osteosarcoma.

Area of Science:

• Oncology
• Computational Pathology
• Medical Imaging

Background:

• Osteosarcoma tumor classification presents challenges due to cellular heterogeneity and data complexity.
• Accurate segmentation and classification of H&E stained histology images are difficult owing to variations and similarities within and between classes.
• Deep learning has shown promise in analyzing cancer histology for breast and prostate cancers.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for enhanced efficiency and accuracy in classifying osteosarcoma tumor subtypes.
• To differentiate between viable tumor, necrosis, and non-tumor tissues within osteosarcoma histology images.
• To provide a computational pipeline for quantifying percentage necrosis in whole slide images.

Main Methods:

• A custom eight-learned-layer CNN architecture was designed, featuring stacked convolutional and max pooling layers for feature extraction.
• Data augmentation strategies were employed to enhance the CNN's performance.
• The proposed CNN was benchmarked against established architectures: AlexNet, LeNet, and VGGNet.

Main Results:

• The proposed CNN achieved an average classification accuracy of 92%.
• The model demonstrated superior performance in distinguishing osteosarcoma tumor classes.
• A functional pipeline for calculating percentage necrosis in whole slide images was successfully developed.

Conclusions:

• Convolutional neural networks offer a viable solution for improving the accuracy and efficiency of osteosarcoma classification.
• The developed CNN model shows potential for assisting pathologists in complex tumor classification tasks.
• The integration of deep learning in digital pathology can significantly advance cancer diagnosis and analysis.