Convolutional neural networks (CNNs) are effective tools for image classification, drawing inspiration from biological visual systems and learning mechanisms. They also offer the advantage of transfer learning, allowing a network trained on one task to be adapted for other potentially unrelated tasks.
A new study from the Oxford University Press shows that scientists can train AI to distinguish brain tumors from healthy tissue.
In a study exploring neural network models, researchers drew a parallel between detecting camouflaged animals and identifying brain tumors. Both tasks involve recognizing patterns that blend in with surrounding environments—camouflaged animals hiding in nature and cancerous cells blending with healthy tissue.
The key to both challenges is the process of generalization, where the network learns to group similar objects under the same identity. The researchers applied this concept by incorporating a unique transfer learning step and training on animal camouflage detection to improve the network’s ability to detect tumors in MRI data.
This retrospective study used public-domain MRI data to investigate how combining these tasks could enhance tumor detection skills in neural networks.
The researchers used MRI data from public online repositories, including sources like Kaggle, the Cancer Imaging Archive, and VA Boston Healthcare System, to train neural networks to differentiate between healthy and cancerous brains, identify cancer-affected areas, and classify cancer types.
The networks performed exceptionally well, achieving near-perfect accuracy in detecting normal brain images with only 1-2 false negatives. One network achieved an average accuracy of 85.99% for detecting brain cancer, while the other had an accuracy of 83.85%.
A key strength of the neural network in this study is its ability to explain its decisions, which enhances trust among medical professionals and patients. Unlike many deep models, which lack transparency, this network can generate images highlighting areas contributing to its tumor-positive or negative classifications.
This feature allows radiologists to cross-validate their decisions with the network’s, providing extra confidence—acting as a second robotic radiologist. The researchers believe that future AI models in clinical settings should prioritize transparency, enabling precise, intuitive explanations of decisions.
While the networks had difficulty distinguishing between different types of brain cancer, they demonstrated distinct internal representations. Accuracy and clarity improved when the researchers incorporated transfer learning, specifically through camouflage detection training, which boosted the network’s ability to identify cancerous areas.
The best-performing model in the study was about 6% less accurate than standard human detection. However, the research successfully demonstrates the quantitative improvements achieved through the training paradigm, particularly with the incorporation of transfer learning.
The researchers argue that this approach and the use of explainability methods are crucial for ensuring transparency in future clinical AI applications. This will pave the way for AI models that are effective but also understandable and trustworthy in medical settings.
The paper’s lead author, Arash Yazdanbakhsh, said, “Advances in AI permit more accurate detection and recognition of patterns. This consequently allows for better imaging-based diagnosis aid and screening, but it also necessitates more explanation for how AI accomplishes the task.”
“Aiming for AI explainability enhances communication between humans and AI in general. This is particularly important between medical professionals and AI designed for medical purposes. Clear and explainable models are better positioned to assist diagnosis, track disease progression, and monitor treatment.”
Journal Reference:
- Faris Rustom, Ezekiel Moroze, Pedram Parva, Haluk Ogmen, Arash Yazdanbakhsh. Deep learning and transfer learning for brain tumor detection and classification. Biology Methods and Protocols. Biology Methods and Protocols. DOI: 10.1093/biomethods/bpae080