Enhancing bone cancer detection through optimized pre trained deep learning models and explainable AI using the...

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate bone cancer diagnosis from histopathology images is crucial but challenging with current computational methods.
• Existing deep learning approaches face limitations in hyperparameter optimization, explainability, and generalizability.
• There is a need for a more reliable and interpretable diagnostic framework for bone cancer.

Purpose of the Study:

• To propose an Optimized Deep Learning Framework for Bone Cancer Detection (ODLF-BCD).
• To enhance diagnostic accuracy, transparency, and generalizability in bone cancer detection.
• To provide a robust computational tool for clinical decision support.

Main Methods:

• Combined Enhanced Bayesian Optimization (EBO) for hyperparameter tuning.
• Utilized deep transfer learning with pre-trained models (EfficientNet-B4, ResNet50, DenseNet121, InceptionV3, VGG16).
• Integrated explainable AI techniques (Grad-CAM, SHAP) for model interpretability and employed data augmentation.

Main Results:

• The EfficientNet-B4 model achieved 97.9% accuracy for binary and 97.3% for multi-class bone cancer classification.
• The framework demonstrated high precision, recall, and F1 scores.
• Explainability methods provided clinical interpretability for model predictions.

Conclusions:

• The ODLF-BCD framework offers a significant improvement over existing methods, enhancing accuracy and transparency.
• The proposed system serves as a reliable alternative to current diagnostic standards like C-RAD.
• This framework can function as a decision support system, aiding clinicians in early and precise bone cancer detection.