Abstract
In research findings, co-deletion of the 1p/19q gene is associated with clinical outcomes in low-grade gliomas. The ability to predict 1p19q status is critical for treatment planning and patient follow-up. This study aims to utilize a specialized MRI-based convolutional neural network (CNN) for brain cancer detection. While transfer learning with models like ResNet and AlexNet is common, these models often contain weights irrelevant to medical imaging, potentially compromising diagnostic reliability. To address this, we developed a custom CNN from the ground up. Our architecture combines convolution stacking with dropout and fully connected layers to reduce overfitting. During training, we augmented the dataset and injected Gaussian noise to improve robustness. Using three-fold cross-validation, the proposed network achieved a 96.37% F1-score, 97.46% precision, and 96.34% recall on a validation set of 156 images, outperforming fine-tuned versions of InceptionV3, VGG16, and MobileNetV2.
Introduction
In neuro-oncology, low-grade gliomas (LGG) are highly sensitive to their genetic composition. Specifically, the co-deletion of the 1p and 19q chromosomal arms serves as a vital biomarker for determining treatment paths and patient prognosis. Currently, identifying this marker often necessitates invasive biopsies. While deep learning offers a non-invasive alternative via MRI analysis, many researchers rely on transfer learning—fine-tuning models trained on general datasets like ImageNet. This can lead to a "trustworthiness" gap, as the models may prioritize features irrelevant to human anatomy. Furthermore, the small and imbalanced nature of medical datasets makes large, general-purpose models prone to overfitting.
Methods
The authors developed a specialized CNN designed specifically for the medical imaging domain, ensuring every parameter contributes to relevant feature extraction.
Architecture
The model features a 14-layer architecture utilizing small 3x3 kernels to capture the fine details of brain tissue edges. It employs LeakyReLU activation functions to prevent the "dying ReLU" problem, ensuring the network continues to learn from negative values.
Training and Regularization
To combat the limitations of a small dataset, Gaussian noise was injected during training. This acts as a form of random data enrichment, preventing the model from memorizing training images and forcing it to learn more robust features. The training process utilized a 3-fold cross-validation strategy to ensure consistent performance across different data subsets.
Results
The proposed network demonstrated high reliability and precision in identifying the 1p/19q marker, significantly outperforming established models:
- Proposed Model: 96.37% F1-score
- InceptionV3: 96.00% F1-score
- MobileNetV2: 92.00% F1-score
- VGG16: 91.23% F1-score
On the validation set (125 codeletion vs. 31 not codeletion images), the custom CNN achieved 97.46% precision and 96.34% recall. Notably, the model achieved a perfect precision score (1.0) for the "not codeletion" class, meaning it produced zero false positives for that category.
Conclusion
This research suggests that for specific medical tasks, a bespoke shallow network can be superior to generic deep architectures. By training from scratch and employing noise injection, the researchers achieved high diagnostic accuracy despite a limited dataset. Future work will explore generative diffusion models for data synthesis, 3D tumor reconstruction, and model compression for deployment on portable medical scanners.
Summary updated