Atif Iqbal, Mohammed Raiyan Nur Hridoy, Ilma Hossain Mim, Riyadul Islam, Siam Tahsin Bhuiyan, Sefatul Wasi, Farhana Sarkar, Saadia Binte Alam
2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN)
In: 2026 IEEE 2nd International Conference on Quantum Photonics, Artificial Intelligence & Networking (QPAIN)
IEEE, pp. 1-6
Glioblastoma (GBM) is recognized as most aggressive and frequently diagnosed malignant brain tumor in the adult population. Its infiltrative pattern and rapid growth makes accurate identification and segmentation of tumor regions crucial for diagnosis and treatment planning. Magnetic resonance imaging (MRI) is a non-invasive medical imaging technique that plays a critical role in the diagnosis of glioblastoma. This study uses MRI scans to train models as they offer high-resolution images of soft tissues, aiding tumor detection. The Brain Tumor Progression Dataset was utilized, which contains the corresponding tumor mask as the ground truth. For the segmentation architecture, we utilized Segformer with the CNN-based models ResNet18, EfficientNet-B3, MobileNet-V2, and DenseNet201 as the encoder. We aimed to implement preprocessing on the annotated tumor masks in such a manner that CNN models can easily recognize the pattern of the tumor masks. It makes the segmentation task more precise during testing. Initially, to train the models, T1 post contrast MRI scans were used, and after training the model with various encoders, EfficientNet-B3 yielded the best results by achieving a precision of 65.85% along with a Recall of 61.68% and a Dice Score of 63.70%. Preprocessing steps were applied to the ground truth masks to remove noisy pixels, fill hollow regions, and reconnect broken portions with the main mask. The models were retrained using the preprocessed masks as ground truth. This led to a higher precision of 79.51% and the highest Dice score of 66.45%, achieved with EfficientNet-B3.