International Journal of Scientific Research and Engineering Development

In the realm of neuro-oncology, diagnosis of a brain tumour using MRI will determine the treatment and the most clinically significant brain tumour types include meningioma, glioma and pituitary tumour. Manual, slice-by-slice radiologic inspection is extremely time consuming, and such inspection is subject to inter-observer variance, for which this deep learning approach is encouraging, particularly by the use of Convolutional Neural Networks (CNNs) that learn discriminative features directly from the raw MRI data. A 23-layer CNN (“Base Model”) is reproduced on the contrast-enhanced MRI dataset of 3,064 images taken on the T1-weighted MRI from the Figshare, trained using Sparse Categorical Cross-Entropy loss and Softmax activation, and the results reported for each class, the confusion matrix and loss/activation ablations are independently computed to ensure consistency with the reproduced CNN. The Base Model has two limitations: 1) It is based on a single fixed preprocessing pipeline applied to all MRI slices, whereas the images can vary in blur, brightness, noise and contrast, and 2) There is no mechanism beyond just scalar metrics to explain the influence on a specific prediction of a given image, which makes visualizing the influence more difficult. The deficiencies it aims to address are tackled by a hybrid pipeline that keeps the validated CNN used and adds a module called Quality Assessment and Adaptive Preprocessing which will learn to do denoising, enhancing the contrast and sharpening on each individual image and a module for explainability named Grad-CAM which will produce a visual heatmap for each prediction. Comparing the proposed hybrid model with the reproduced Base Model, a higher accuracy (98.04%), average precision (97.80%), average recall (97.95%) and average F1-score (0.979) are achieved on the same stratified (70/15/15) dataset of the same Figshare data. The results demonstrate the robustness and interpretability of existing CNN models can be increased with only per-image adaptive preprocessing and increasing the interpretability using explainability provided by Grad-CAM, and there is still an important future direction to validate the model with an external dataset, and to present the explanations to radiologists.

Jyoti, Amandeep, Dharmender Kumar, "Brain Tumor Detection Using CNN" International Journal of Scientific Research and Engineering Development, V9(4): Page(541-548) May-June 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.