Background: The most common type of cancerous bone tumour is Osteosarcoma. Microscopic analysis of coloured slides is mostly used for Osteosarcoma diagnosis. As Osteosarcoma is highly heterogeneous, pathologists struggle to get a consensus diagnosis. Moreover, manual analysis of microscopic images is laborious, time-consuming, and vulnerable to bias.
Method: In this research, a framework that extracts features based on CNN is developed to predict Osteosarcoma to address previous studies' performance gaps and shortcomings. The dataset used for this experiment is obtained from pathology archives at Children's Medical Center in Dallas, which includes 1144 histology glass slides. First, data augmentation is used to extend the dataset size. This study found Blur, Contrast, and Random augmentation perform well for the Osteosarcoma dataset. Next, we primarily considered six pre-trained transfer learning (TL) models and modified the models to extract features. Then, we examined feature selectors like Genetic Algorithm, Principal Component Analysis, Recursive Feature Elimination and Baruto and found PCA selects the optimal feature set. Finally, we fed the selected features into a fine-tuned MLP called Grid search MLP (GsMLP) and utilised federated learning (FL) to ensure the diversity and privacy of the dataset.
Results: Experiments have revealed that the modified Xception model performs better in extracting features for Osteosarcoma prediction. Also, the proposed framework performed better than state-of-the-art techniques such as Convolutional Neural Network, VGG19, Random Forest, and AdaBoost with 98.20% accuracy and 97.5% Receiver Operating Characteristic - Area Under the Curve.