False Negative Induction in Brain Tumor Segmentation by Trained Noise Attack

Diagnosis of medical conditions more specifically brain tumors requires high accuracy; any alteration to that due to system glitches may result in detrimental consequences. U-Net’s architecture is found suitable for semantic segmentation tasks like brain tumor detection for its efficient feature extraction, ability to handle limited training data, and state-of-the-art performance. In this work, we investigated if the insertion of a trained noise in brain scan images can induce false negatives in a white box arrangement (where the information about the model is known). Our investigation on the insertion of a noise trained using the gradient descent method is found to generate false negative errors in the prediction of U-Net architecture. The contribution of this work includes a comprehensive method to generate trained noise that can produce false negatives when inserted into a brain scan image. To our best knowledge, there is no such work of noise insertion attack that has put forward the vulnerability of popular U-Net architecture in medical image segmentation, especially in brain tumor detection. This work aims to highlight the vulnerability of automatic brain tumor detection using U-Net architecture under the attack of trained noise and to maneuver a future scope of research based on this investigation.